Full Archon Guide - Build AI Coding Harnesses That Actually Ship (LIVE)

0:04All right, we are live.

0:07Welcome everybody to the Archon live

0:11stream. We're going to be diving into

0:14everything with Archon today. So I

0:16posted a YouTube video just this week

0:19introducing the new version of Archon,

0:22but now we get to actually go deep into

0:25using it and there are a couple of

0:27things I have up my sleeve today that

0:29are kind of interesting. So I'm going to

0:32be

0:33building a what's called a dark factory

0:35this month. If that sounds spooky or

0:37interesting to you, we'll talk about

0:39that more in a little bit.

0:41That'll be kind of like the tail end of

0:44our live stream today.

0:46So yeah, but I want to start by actually

0:48introducing Archon to you guys just to

0:51give a whole, you know, overview of the

0:54tool. I know that I talked about it on

0:56Wednesday in the YouTube video, but I I

0:58want to be pretty comprehensive in the

1:00live stream today and then we'll build

1:02some workflows together. I'll show you

1:03what it looks like to use Archon, how I

1:04use it on a daily basis and then we'll

1:07get into the the dark factory stuff.

1:10I'm pretty excited. So

1:12yeah, I'm sitting down right now. I'm

1:14like actually like always standing in my

1:16live streams and my YouTube videos, but

1:19I had a couple of leg days like in a row

1:24and my legs are like toasted right now.

1:26So it's really it actually kind of like

1:27hurts to stand right now.

1:29Pretty crazy. So yeah, I'm actually

1:32sitting down for a live stream. I've

1:33like never done that before.

1:35But yeah, anyway, so I've got my left

1:37monitor here with all of the comments

1:40from you guys that I'm watching and then

1:42I've got my right monitor where I have

1:44my streaming software. So when you see

1:46me look around, that's what I'm doing

1:47and then obviously I'm sharing my screen

1:50on the main monitor right here.

1:52So I can actually switch to share my

1:54screen here and then I can pop up chats

1:57like this right here. Hey Cole, looking

1:59forward to this one. I appreciate it

2:00very much. Yeah, so I was I was checking

2:02out the chat as I was getting things

2:04started here. It's it's exciting to see

2:06everyone in here already. Congrats on

2:08the launch. Love seeing the structure

2:10for AI workflow. Still wrapping my head

2:12around its potential. You know,

2:14I appreciate Thomas and I still am as

2:16well cuz it's actually kind of crazy.

2:20Maybe we'll talk about this a little bit

2:21today, but we can use Archon for a lot

2:23more than just AI coding. Like you can

2:26use it for any kind of agentic workflow

2:28for deep research or any kind of like

2:31content creation.

2:33I'm not going to build anything around

2:35that now, but that's going to be more

2:37coming soon in videos and future live

2:39streams. So yeah, the the possibilities

2:42are pretty limitless here and I think as

2:44I really introduce Archon to you guys

2:46right now, you'll see what I'm talking

2:48about.

2:50Cool. I'm good sitting down. You still

2:52look great, Cole. Appreciate it, Eric.

2:54Yeah, it feels weird to me, but

2:57yeah, I definitely need it right now.

2:59It's exactly right. Exactly right. Never

3:00skip leg day.

3:02Yeah, so so I started doing so I was

3:05doing a lot of CrossFit last year

3:09and I took a break so I just like

3:11started going to like a regular

3:13commercial gym and I and I went back to

3:15doing CrossFit again

3:17and the workouts were pretty intense

3:20last couple days for the legs. So

3:22yesterday was a lot of Bulgarian split

3:25squats, jumping lunges, wall balls and

3:29um

3:30and biking.

3:32So yeah, I'm pretty cooked, but it's

3:34good. [laughter]

3:36All right. Cool. So

3:39I'm going to go ahead and

3:42pop up my Obsidian vault here.

3:46So I've got the diagram to really

3:48introduce

3:49why I built Archon in the first place.

3:53And I also want to be clear here. I did

3:56not build Archon by myself. There are a

3:59couple of people that helped me a lot

4:02with Archon.

4:03Rasmus and Thomas are the two that have

4:07really helped me the most with the

4:08project.

4:10Thomas is here. He's a DIY smart code.

4:13So he's here. I see him in the chat.

4:15Shout out to Thomas. Not sure why

4:16there's an echo for the stream, Thomas,

4:18but yeah, thank you for all the work

4:20that you've done on Archon. Rasmus, I

4:21don't know if he's here. But yeah, big

4:24shout out to him as well.

4:27Yeah, I don't think he is for today, but

4:30that's all good. So yeah, I've I've had

4:33I've had a lot of help building Archon

4:34and of course I've had a lot of help

4:36from Claude Code at building Archon as

4:38well.

4:39Let me tell you, I am using Archon to

4:42build Archon constantly. I'll show you

4:45guys what that looks like today as well,

4:46but let's talk about the evolution here.

4:49Why we care about Archon. Why why did I

4:51care about building it in the first

4:52place? So allow me to like talk about

4:55this for a little bit

4:56cuz I really want to get you guys on

4:58board with the vision that I have here

5:00and it is a big vision.

5:02So

5:03it's everyone knows that like prompt

5:05engineering was a big thing when

5:08generative AI first became a thing back

5:11in 2022. So we had like, you know, the

5:13release of GPT 3.5 Turbo. That's when

5:17everyone started giving a lot of hoots

5:19and hollers about generative AI and

5:21people were obsessed with prompt

5:23engineering. And the idea behind prompt

5:25engineering is how can we craft our

5:28prompts to the LLM to get the single

5:30best output. So at this point people are

5:33very much like caring about just you

5:35know, that very next turn with the LLM.

5:38How do we get it to do or output what we

5:40want?

5:42And then that evolved in 2025 to the

5:46idea of context engineering. So now we

5:48don't care about just single outputs.

5:50Now we care about entire sessions,

5:52especially for AI coding sessions.

5:54Context engineering is the idea of you

5:57know, how can we curate the perfect

5:59context. Nothing more than the coding

6:02agent needs, but exactly what it needs

6:04to have all the context it needs to

6:07plausibly do the task.

6:09So this was popularized by a lot of

6:11people like Toby, the CEO of Shopify. Of

6:14course Andrej Karpathy cuz whenever you

6:16post something, of course it goes viral.

6:19But context engineering, it's a very

6:20powerful evolution because it's treating

6:24the context for our coding agents as an

6:27engineered resource. So we evolve our AI

6:31layer as I like to call it just like we

6:33evolve our code base. So we version

6:35control our rules and our commands and

6:37our skills and whenever there's a

6:40problem that comes up with our

6:42system, like there's a bug that our

6:44coding agent produces, instead of just

6:46fixing the bug and moving on, the idea

6:49behind context engineering is we look

6:51into our process. Like what could we

6:53make better with our commands or what

6:55could we make better with our rules so

6:56that that issue doesn't happen again. Or

6:59at least that's the goal of evolving our

7:01context.

7:02So it's a very powerful concept and I

7:04created a lot of content last year

7:06around context engineering. Now the

7:08important thing is these different

7:10evolutions, it's not like they replace

7:12each other. Context engineering does not

7:15replace prompt engineering. In fact,

7:17prompt engineering is a part of context

7:19engineering. So it's it's a evolution

7:22where it builds on top of itself.

7:24And so this year we have evolved to

7:27harness engineering. And this is what

7:29Archon is all about. So harness

7:32engineering can mean a lot of different

7:34things, but in essence, a harness is a

7:38layer on top of the coding agent. It's

7:41the tooling and the process that you

7:43build on top of the coding agent to make

7:46it more reliable and basically taking

7:49your whole you know, process for

7:52building software

7:54and you know, building the the layer on

7:56top of the coding agent where you

7:57enforce that process. So a lot of

8:02different strategies out there like B

8:03Mad and GitHub Spec Kit and GSD, you

8:07could consider them harnesses because

8:09you're you're wrapping the coding agent

8:10with this higher level layer of your

8:13strategy for engineering. And we have a

8:16Stripe Minions. The Ralph Loop is a

8:18harness. Anthropic has open sourced

8:21quite a few harnesses. I've covered some

8:22on my channel.

8:24Really every single big company is

8:27converging on the idea of harnesses.

8:30Instead of focusing on making the model

8:32better or making the tool better like

8:34the coding agent better, it's it's all

8:35about like how do we create the system

8:37that wraps the coding agent, right?

8:40And so there's a lot of studies that

8:42have been done that have shown like even

8:44if you are using the exact same model,

8:46you're not improving underlying model or

8:48tool, you can go from a 6.7% pull

8:50request acceptance rate to almost 70%

8:54and the only thing is the harness. So

8:56building in your strategies around

8:58context curation and validation and your

9:01approach for planning, right? Like we'll

9:03talk about all that today when we cover

9:05building harnesses with Archon, but it's

9:07a big deal. Like maybe you guys heard of

9:09Stripe Minions. This went like super

9:11viral last month because Stripe

9:13they built their own internal harness.

9:15It's kind of like a more powerful

9:17version of the Ralph Loop that allows

9:19them to ship 1300 AI only generated pull

9:22requests every single week. And like

9:25Claude Code itself is starting to build

9:27a lot of harness into the tool. So not

9:31just calling the coding agent, you know,

9:34calling the LLM with the coding agent,

9:35but how can we like orchestrate many

9:37different Claude Code sessions together?

9:40So like Claude Code has support for

9:42agent teams. They're doing a ton of sub

9:43agents right now. So when they had their

9:46source code leak last month,

9:48it was shown that like 60% of the code

9:51is still like wrapping the underlying

9:53model itself like Sonnet and Opus and

9:55Haiku,

9:56but then 40% of their code is with these

9:59more like harness features like agent

10:01teams. And so all that to say, it's the

10:04direction that all of these companies

10:06are heading where they're building

10:08harnesses. The coding tools themselves

10:10are building this. All of the biggest

10:12companies like uh Stripe and Shopify and

10:15AWS, they're all building their own

10:16internal harnesses.

10:18But the problem with all these harnesses

10:19is they're not open-source and they're

10:21not custom to you.

10:23What if you want to take your software

10:26development life cycle, your process for

10:29working with AI coding assistance, and

10:31what if you want to package it up into

10:32your own harness? Well, that is the

10:35value proposition of Arkon. Arkon is the

10:38harness builder. So for the first time

10:40ever, we have an open-source platform

10:43that makes it easy for you to build your

10:45own harness. Up until this point, you've

10:48either had to build something internally

10:50like Stripe did, but obviously that's a

10:52massive amount of effort, or you just

10:55had to use a harness that's already out

10:57there like the Ralph Loop or B mad or

11:00whatever it is. And like yeah, those

11:01tools are very powerful, but they're not

11:03custom to you.

11:05And you you guys know if you've been

11:07following my content that uh I'm always

11:10a big proponent of like build it

11:12yourself. Take inspiration from what's

11:14already out there. And I'll even show

11:16you what it looks like to build

11:17something like a B mad or GSD harness in

11:20Arkon cuz we can take inspiration from

11:22the beautiful minds that are out there,

11:24but still make something that uh works

11:26how we want to work. And that's the

11:28unlock that I have for you guys here

11:30with Arkon.

11:32And um the other thing that I want to

11:34cover here is

11:36um you know, everyone already has their

11:38own skills and commands and rules. So

11:41I'm not I'm not expecting you to like

11:43start from zero. I know that you already

11:45have a process that you built around

11:48your AI coding assistant. Like you have

11:50skills for validation. You have commands

11:52for planning and maybe you have your PRD

11:55template, whatever that might be. And so

11:57it's not like Arkon replaces what you

12:00already do. It's that Arkon allows you

12:03to package everything together into a

12:06workflow that combines your skills and

12:09commands, right? Like the whole idea is

12:11before you have a harness, you have your

12:15commands and skills and you have to

12:16remember the order that you use them.

12:19You have to work between different AI

12:20coding sessions because you definitely

12:23don't want to do your planning in the

12:24same session where you do your

12:25implementation cuz your coding agent

12:27builds up a lot of bias.

12:29But if you don't have a harness to

12:31connect those steps together, you have

12:33to be the one to orchestrate that. I

12:35call it shepherding, right? Like you you

12:37go through some kind of PRD process with

12:40your coding agent. You create a PRD and

12:42then you're like, okay, good. Now let me

12:43go over to a new Claude code session or

12:45a new Codex session or whatever, and let

12:48me go and create a plan for the first

12:50phase. And then you get your plan and

12:52then you go and you create a new session

12:53and you go into implementation. And then

12:56you have a new session where you do a

12:57review on the pull request or on the

12:58code, whatever it is. And so you're

13:00still, even though you have your

13:02commands and skills and things to

13:04automate different parts of your

13:05workflow, you still have to walk the

13:07different coding agent sessions through

13:09each one. And Arkon is kind of like the

13:12next evolution of that. It allows you to

13:16run longer tasks, but still keep

13:18yourself in the loop. And so I have an

13:21example here

13:22of just what an Arkon workflow can look

13:25like. And and by the way, if you if you

13:28did watch the video on Wednesday, I am

13:31repeating myself a little bit here. So I

13:32covered this diagram in the video, but

13:34uh you just consider this a refresher if

13:36you watched it. And I just want to make

13:37sure that for those of you tuning into

13:39this live stream who didn't watch the

13:41video, I can like give you a really

13:42solid overview of what Arkon is. Um so

13:46yeah, here's just one example of the

13:47kind of workflow you can build with

13:49Arkon. And uh you'll have to excuse that

13:51the images are like not rendering right

13:53now for some reason for some of the

13:54icons, but my point can still be made

13:57here. Uh but yeah, so anyway, with

13:59Arkon, you obviously have some kind of

14:01trigger for your workflow.

14:03So traditionally without a harness, it

14:05would just be like sending a prompt into

14:07Claude code, right? Like that is your

14:08trigger. But with Arkon, there's a quite

14:10a few different ways that we can use it.

14:12And I'll I'll show you guys that today.

14:13We have the CLI. So we can have our

14:16Claude code or Codex, whatever um

14:19instance like trigger an Arkon workflow.

14:21So we can dispatch work to delegate

14:24behind the scenes. So we can do a lot of

14:26work in parallel. I also have a whole

14:28web interface. And I I'll definitely

14:31show that off today as well.

14:33Um that's the wrong link. This is it. So

14:35we have a whole web interface where we

14:36can manage our workflows

14:39and we can uh kick off workflows

14:42uh from the web UI directly.

14:45And uh so there's basically like a a

14:47coding agent sitting behind the web

14:50interface. So we can ask it to do work

14:52on our different repositories that we

14:54have registered with Arkon and it'll

14:57automatically route the request. It'll

14:59pick the right workflow and it will

15:01dispatch it. And we can do a ton of

15:03workflows in parallel if you want as

15:05well. So I'll I'll show you all of that

15:06today. But we have the web UI and then

15:08we even have different adapters so that

15:10you can, for example, talk to Arkon

15:12directly in Slack. So you can say like,

15:14hey, there's a GitHub issue for um this

15:17repo. Go and you know, handle that

15:20issue. I want to see a pull request in

15:21the end with a fix for that issue. And

15:23you can do that through Slack. And we

15:25have different conversations that you

15:26can manage through different threads in

15:28Slack. So for every single platform, we

15:31support parallel execution.

15:33Arkon handles work trees and isolation

15:36under the hood so you don't even have to

15:38worry about it. That's one of the other

15:40big unlocks with Arkon is how it

15:42supports working on things in parallel.

15:44And so I can go into my Claude code and

15:47I can have it use the CLI to fix eight

15:50GitHub issues at the exact same time.

15:52And they all run in different work trees

15:54so they don't step on each other's toes.

15:56They don't override each other's changes

15:57and we don't have to deal with merge

15:58conflicts just to get the pull requests

16:01created. It's really powerful.

16:03And so we start with our trigger. And

16:05then um again, this is just an example

16:07of a workflow we can create. So we could

16:08go like into planning mode. So maybe

16:11there's a new feature that we want to

16:12build. We want to start by going through

16:14the planning process with our coding

16:16agent.

16:17And so we have a step in our Arkon

16:21workflow where we are prompting our

16:23coding agent like here is the feature we

16:25want to build. Now help me plan. Like do

16:27research, do code base analysis,

16:29whatever that is. And when in Arkon, we

16:32can add in human in the loop. And so we

16:34can have essentially a loop here where

16:37we give feedback to have the coding

16:38agent revise the plan until we approve

16:41it and then it moves into the coding

16:43stage. And we do this in a brand new

16:44coding agent session because your

16:47planning session can get pretty bogged

16:48down and you can build up a lot of bias

16:51over time. And so you want to produce an

16:54artifact that you send into the next

16:56node for implementation.

16:59If you're in the Dynamis community, this

17:00is the pivot loop. You know what I'm

17:02talking about. I've covered it on my

17:03channel as well. The plan, implement,

17:05validate. And we used to have to do that

17:07between different coding agent sessions,

17:09but now we can package this all up as a

17:11single Arkon workflow.

17:13And then after the code is done, then we

17:15go into the testing step. And this is

17:18another one of the very big value

17:20propositions of Arkon. Our workflows do

17:24not have to be just prompts to a coding

17:27agent. Sometimes there are steps that we

17:30want to run deterministically. We want

17:32to take the control away from the coding

17:35agent to make sure that our process is

17:38followed to a T.

17:40If you've used any coding agent for a

17:42good amount of time, you know that like

17:44sometimes you'll tell the coding agent

17:46run the tests after you write the code

17:49and it won't listen to that. It's so

17:51frustrating. Or it'll do some of the

17:53testing like unit testing and linting,

17:55but it won't do the end-to-end testing.

17:56It's so frustrating. And so what we can

17:59do in Arkon is we can have certain steps

18:01of the workflow where we're just running

18:03code. Like we're doing some kind of

18:05context pulling or we're running our

18:08tests like we're doing right here. We're

18:10guaranteeing that that happens after the

18:12implementation. And then if there's any

18:14failures, we'll prompt the coding agent

18:15to fix those things. So we have like a

18:17little bit of a feedback loop as well

18:20built directly into Arkon.

18:23And then after everything passes and the

18:25coding agent figures that it's done,

18:27then we have the human approval.

18:29And so even though we are building

18:32longer running tasks with Arkon

18:34workflows, we're not taking ourselves

18:36out of the loop. We can inject ourselves

18:38wherever we want in an Arkon workflow.

18:40So no matter how you typically work with

18:42your coding agents, if you don't trust

18:44it that much and you want to validate it

18:46every step of the way, you can do that

18:49if you want. We have support built in to

18:51the web UI, the CLI, every single

18:54adapter for human in the loop. And

18:56that's one of the most important things

18:57because a lot of times when you have

18:59these harnesses, especially like the

19:01Ralph Loop for example. The Ralph Loop

19:03went viral a couple of months ago, but

19:05to me it felt like vibe coding, right?

19:08Because you were giving it a and it

19:09basically an entire PRD. Like many

19:12different phases of work and you would

19:14just have a coding agent rip through

19:16everything.

19:17And the problem with that is if the

19:19coding agent makes a mistake in the

19:21first iteration of the Ralph Loop, that

19:23issue can kind of propagate and like

19:25blow up from the rest of the loop

19:27because the coding agent is going to

19:29through each loop build on top of a code

19:31base that's already wrong, right? Like

19:33not aligned with what you actually want

19:35to create. And so the issues just

19:37compound on themselves. And so the idea

19:40that I wanted to be uh very confident in

19:43for Arkon is that it's not just enabling

19:46vibe coding. That we actually have like

19:48deterministic steps that enforce our

19:50process. We have human in the loop,

19:52right? And then we have the pull request

19:54at the end, and then we can review that

19:55our self before we merge as well.

19:58And so, yeah, that's that's sort of I I

20:00like to call it the hybrid secret for

20:02Archon.

20:03And um this this sometimes rubs people

20:06the wrong way, but I actually think it's

20:08smart to take as many decisions from

20:11away from the coding agent as you

20:12possibly can because they're

20:14non-deterministic. They don't make the

20:17same decisions every time even if you

20:19give it the same prompt. And so, there's

20:21a lot of inherent risk with that. The

20:23biggest reason why a lot of developers

20:26and companies are hesitant to adopt

20:29coding agents is because they're

20:31unreliable by nature. And so, Archon is

20:34the harness builder that allows you to

20:37take the process that you consider

20:39reliable and build it into the way you

20:42work with coding agents. And so, we have

20:44different nodes. Like if you have your

20:46testing strategy, if you have your

20:48strategy for like pulling information

20:50from Confluence, like whatever steps you

20:51have that like you want to have

20:53performed every single time, you build

20:55those nodes into Archon.

20:57And then the rest of the workflow is

20:58still going to be like you adding in

21:00your commands and skills and sub agents.

21:02Like everything that you already have,

21:03you can build in. So, you're not losing

21:05anything either. That's another thing

21:06that I want to be very clear on here is

21:09that you don't have to replace how you

21:11already work. Archon just allows you to

21:13package it up, right? Like that's the

21:15goal of a harness builder. So, I'm I'm

21:18so excited for this. And I'll I'll

21:21actually like install it from scratch

21:22with you guys today as well. So, uh

21:25right here I got the link to Archon. I'm

21:27going to go ahead and uh drop this in

21:28the chat.

21:31So, if you guys want to

21:34um try it out right now, even like

21:37install it along with me, I would highly

21:40encourage you to. It's very easy to

21:42install Archon because we can ask our

21:46coding agent to set it up. And I have an

21:48Archon skill that it loads

21:50automatically, and it'll walk you

21:52through the entire process, even

21:54installing the dependencies for you. So,

21:56I'll show you that live in a sec, but

21:58yeah, it's so easy to get it installed.

22:01Um yeah, so one other thing I want to

22:04show really quick. So, this is the read

22:05me for the new version of Archon.

22:08I want to show what a workflow actually

22:10looks like.

22:11So, this might look a little bit

22:12intimidating,

22:14but uh don't worry.

22:16You can use your coding agent to help

22:18you build workflows as well. I'll also

22:20show you that.

22:22Um so, this whole like YAML structure,

22:25it's it's quite beautifully simple,

22:27honestly.

22:28And uh but like even even so, like when

22:31you're trying to package up your entire

22:32workflow, it can start to look a little

22:33intimidating. So, it's really nice to

22:35use a coding agent to help you define

22:38these workflows.

22:40But just an example here. So, within the

22:42diagram here, I showed the whole like

22:44PIV loop workflow.

22:46And that that workflow can be summed up

22:49with this right here. So, this is a kind

22:50of a bit of a simplification just so it

22:52fits nicely in the read me, but it gives

22:54a good idea here. So, you start with the

22:56node for your planning, and then you go

22:58into implementation. And you can see

23:00that we have we support loops in Archon

23:02as well. So, uh maybe I can even go into

23:05the UI. Let's see if I have it run with

23:08a loop.

23:09I want to give like a demo of like what

23:11it actually looks like. Yeah, here we

23:12go.

23:13So, like we have we support loops in

23:15Archon. So, by the way, this is what the

23:17logs look like. When you invoke a

23:19workflow and you view the logs in the

23:20web UI, you can see like all the tool

23:22calls that your coding agent is making

23:23behind the scenes. You can see like the

23:25the node flow and like where it's

23:27currently running. So, like we have a

23:29loop right here that has like a certain

23:31number of iterations. And uh if I go to

23:34like a more fancy one like a GitHub

23:35issue fix, we can see like you can get

23:38pretty comprehensive with the workflows

23:39that you define. This is one of the the

23:41bread and butter workflows in Archon, uh

23:45the fix GitHub issue. So, you can throw

23:47this on any GitHub issue in any

23:49repository, and it works through this

23:52entire process of uh you know,

23:54classifying the issue. Is this a a new

23:57feature we have to plan, or is it a bug

24:00we have to investigate?

24:01And then it'll do some research. It'll

24:03fix the issues after investigating, and

24:06then it will validate, and then create a

24:08pull request and review it after. So,

24:10it's like super comprehensive workflow.

24:12And and yes, having this many steps does

24:14take a good amount of tokens, but

24:16another really powerful thing you can do

24:18in Archon is specify the model the at

24:21the individual node level. So, when

24:23you're classifying or investigating,

24:25maybe you only need to use Haiku. And

24:27so, it might only be the case that like

24:29a single node like the fix issue, this

24:31is the only one where you'd want to use

24:32Opus. Or you'd want to use high

24:34reasoning GPT Codex, for example. And

24:37so, you can make things very token

24:39efficient by determining at at each

24:41individual level what model do we want

24:43to use.

24:45And the other really powerful thing

24:46about harnesses, like I showed in this

24:49diagram here, is you get insane results

24:52building a harness on top of a model.

24:54So, the harness elevates the model. I've

24:56had better results using Archon with

24:58Sonnet than I have using Opus by itself

25:01in Claude Code.

25:03And so, uh another really cool thing is

25:05like

25:06you know, this um Archon kind of comes

25:09with good timing here because Anthropic

25:11has made their rate limits a lot worse

25:13for Claude Code recently. It's really

25:15unfortunate. But it's forced me to like

25:17start using Sonnet more for my coding

25:20because I hit my rate limit so

25:21incredibly fast. But with Archon, I feel

25:24like I'm not really losing the code

25:27quality, like the output quality,

25:30because the workflows

25:32they package up such a comprehensive

25:34process. And like using this whole thing

25:36with Sonnet is still cheaper than like

25:38asking Claude Code to fix an issue by

25:40itself with Opus cuz Opus is just so

25:42much more expensive.

25:44So, anyway, I got off track a little bit

25:46showing you guys the web UI here, but

25:48yeah, so we have the loop like we can,

25:50you know, implement a plan task by task

25:52in a loop, and then we run the

25:53validation. And the important thing is

25:55this is no AI. It is deterministic. So,

25:57we're guaranteed that our validation

26:00runs. That's powerful. And then we run

26:02the reviews afterwards, and then we have

26:03some kind of approval process with a

26:06human in the loop before we create our

26:08final pull request.

26:10And because we have Archon as a CLI and

26:13a skill, when you have it loaded into

26:16your Claude Code or your Codex, all you

26:19have to do is tell it to use Archon.

26:22And that like literally that's all your

26:23request has to be. And it'll

26:25automatically dispatch like create an

26:27Archon workflow. So, it'll run like

26:29let's say for example this idea to PR

26:31workflow. It'll create the work tree

26:33under the hood so you can keep working

26:34on your code base in parallel. And then

26:36it goes through the different stages,

26:38and you can have your coding agent

26:40monitor the workflow and or monitor it

26:43in the web UI like I showed right here.

26:45So, like obviously this is a completed

26:47run right now.

26:48But uh if this was in the middle of

26:51executing, you could like see the logs

26:52coming in real time and see like, you

26:54know, like what stage of the workflow

26:56the coding agent is currently on.

27:00And as far as which coding agents we

27:02support right now, we have support for

27:05Claude Code, and then Codex is almost

27:08done.

27:09And then also, we are working on uh or

27:13we want to add in support for other

27:14coding agents as well. So, like right

27:16now it's Claude and Codex,

27:18but uh we have a priority item right now

27:20to add support for Pi as well. So, Pi

27:23also has an SDK. Uh we're also

27:26interested in Amp and Open Code. I mean

27:28really like Archon, remember Archon is

27:30the layer above the coding agent cuz

27:32it's the harness builder, right? So, we

27:34don't really care about what coding

27:36agent you're using under the hood. We

27:38just have to build a little bit of

27:39support for like calling the SDK for the

27:43coding agent. So, like Codex has the

27:45SDK, Claude has the agent SDK. That's

27:48what Archon uses under the hood. So,

27:50when you use Claude with Archon, it is

27:52running Claude Code. It's just running

27:53it programmatically through the SDK.

27:57And um so, you can you are allowed to

28:00use your Anthropic and Codex

28:02subscriptions with Archon. You don't

28:05have to pay for API credits.

28:08Um and and that's that's a very

28:09important thing cuz a lot of people

28:12right now are getting their Anthropic

28:13subscriptions banned when they use it

28:16with Open Claw and Open Code.

28:19But the problem with those they're

28:21third-party harnesses. They're kind of

28:23doing a workaround to use the

28:25subscription, and that's against the

28:27Anthropic terms of service. But

28:29Anthropic has made it very clear. A

28:31couple of their team members like posted

28:33on X saying that like you are allowed to

28:36use your Anthropic subscription

28:38with the Claude agent SDK as long as it

28:42is for personal use. And Archon is for

28:44personal use, right? Like you're running

28:46it on your computer, you're hosting it

28:48yourself, and there aren't other people

28:51using your subscrip- subscription

28:53through a production deployed agent. So,

28:55it's a different story if other people

28:57are using your subscription through a

28:59Claude agent SDK agent you have like

29:01deployed to some production platform,

29:04then that's against the terms of

29:05service. But if it's a an application

29:06you're running yourself and is using the

29:09Claude agent SDK, then you are allowed.

29:11So, I've been using my subscription with

29:14Archon for, you know, months and months

29:16now. And the same thing with my second

29:18brain as well.

29:19And so, if they've made that very very

29:21clear. I don't have the exact tweet up

29:23right now, so I can't show you the exact

29:25thing, but Boris Cherney, he's the

29:27creator of Claude Code.

29:29He um he clarified like he made it super

29:31clear. Like yes, you can use your

29:34subscription for personal use with the

29:36agent SDK. So, we're good.

29:38>> [laughter]

29:38>> And I get that I get asked that question

29:40all the time as I'm doing my second

29:42brain content on YouTube, and I'm you

29:44know, I started showcasing Archon. And

29:45it is a fair question because um yeah,

29:48like people are really nervous about

29:49getting their subscription banned, but

29:52you are good with Archon.

29:54And yeah, I see some things in the chat

29:56here. People are excited for the pie

29:58support. Yeah, I'm excited for that as

30:00well.

30:01Because then it you will have like

30:02support for pretty much running any

30:04model that you want with Archon.

30:08All right.

30:10Cool.

30:11All right. So yeah, I'm going to go

30:13ahead and answer some questions in the

30:15chat here, and then we'll get right into

30:18installing Archon. So I have an instance

30:22already spun up

30:23in the cloud. I I installed Archon from

30:26scratch on my computer for the video on

30:28Wednesday. I don't really want to like

30:29reinstall it on my computer, cuz I

30:31already have some things going with it.

30:32So I'm going to reinstall it on a Linux

30:35VPS that I created just in Digital

30:37Ocean. But the installation process is

30:39going to be the same pretty much no

30:41matter your operating system. So whether

30:42you're installing it on a VPS or on your

30:45Linux, Mac, or Windows system, it's

30:46pretty much going to be the same.

30:48And the readme has good instructions for

30:50that. And I'll I'll walk you through it

30:51right now as well. But yeah, before we

30:53do that, I'm going to switch here to my

30:56full frame. Let's let's chat a little

30:59bit. So I'll I'll pop up some questions

31:01in the chat here

31:03and spend some time for Q&A.

31:08All right.

31:09Jared said you said the magic words,

31:11customize B for inspiration into custom

31:13workflow. That's exactly right. So it

31:16works the way you need it, just the way

31:18I need it. This is fire, what I've been

31:20looking for.

31:21Yeah, I appreciate it a lot.

31:24And that exactly, like that's that's the

31:25thing is

31:26you can just like

31:28fork B mad, and you can just run it and

31:31change it yourself, but that's a lot

31:33more involved than creating an Archon

31:34workflow. I'll show you in a little bit

31:36what it looks like to customize

31:38a workflow taking inspiration from

31:41something like B mad. I'm going to use

31:42GSD as an example, cuz it's a bit

31:45simpler and easier to do, but you could

31:46do the same thing with B mad.

31:51All right.

31:55So basically, we could call Archon

31:57Stripe Minions, but open source.

32:00Exactly. Yep. And I'd even go further to

32:02say that like Stripe Minions is just a

32:04single harness. Archon can allow you to

32:06build anything. Like Archon allows you

32:09to build Stripe Minions, but for your

32:11company or yourself.

32:13Yeah.

32:15Can a node be simply an execution node,

32:19no AI model used? Yep. Yep. So we have

32:21support for running bash, Python, and

32:24TypeScript scripts. So anything like any

32:27part of the workflow where you're like,

32:28I want this thing to happen exactly,

32:30like no AI model to mess it up, then you

32:33can run it as a node in Archon.

32:35Yeah.

32:37Where does this differ from N8N?

32:40You know, that's a really good question.

32:42Actually, here, I'm going to switch back

32:44to my scene here. I'm going to share my

32:45screen. Cuz if we look at the the

32:48builder here, and sorry, I'm going to

32:50the wrong tabs. We're actually adding

32:51support soon for a visual builder. So

32:54you can like connect the nodes like

32:56this.

32:57And so it is going to be kind of similar

32:59to N8N. In fact, in the readme here,

33:02we say think N8N, but for software

33:04development. And so for those of you who

33:06use N8N or have used it in the past,

33:08this might actually click really nicely

33:09for you. Is like N8N doesn't really have

33:12direct connections with Claude code. And

33:15they don't have things like work tree

33:16isolation support, right? Like you don't

33:18use N8N for AI coding. You use it to

33:20build automations for your business,

33:22whether it's like LLM workflows or not.

33:25But Archon is going to allow you to

33:28build kind of like N8N workflows, but

33:32specifically for AI coding. So you're

33:34stringing together these coding agent

33:37sessions and deterministic steps like

33:39bash and Python scripts, and it's for

33:41coding processes. And so like there's a

33:44lot of things we have built in behind

33:45the scenes with Archon for how it

33:47integrates with our different coding

33:49agents and handles work trees and things

33:51like that that N8N doesn't have at all.

33:53So like there's no way you could build

33:55this kind of thing in N8N. You're you're

33:58not going to be able to build a coding

33:59agent harness in N8N, or at least it'd

34:00be very difficult and not really what

34:02the tool is meant for. So this is a very

34:05different use case, but you can think of

34:07it like N8N for AI coding. So yeah, I

34:09appreciate you asking that.

34:13All right.

34:15How can we know that the plan step is

34:18complete? Do we have human and approval

34:20gate for that?

34:22Yes. So you can have the coding agent

34:25decide itself that the plan is complete,

34:27or you can make it so that you have to

34:29approve.

34:31And so like going back to the diagram

34:33here,

34:35this step right here is what you're

34:36talking about. Like planning. And so

34:39what it does is it'll output some

34:40artifact. And you can have I know that I

34:43only show the human approval gate here,

34:45but you could like add that as a node

34:46right here as well. So like you get to

34:48look at the artifact, and it'll like

34:50give you a summary of it, or it'll give

34:52you like the full files, so you can like

34:54go through the markdown if you want.

34:56And then you can give feedback. So you

34:58can say like, "Hey, you didn't do enough

35:00validation, right?" So or like you

35:02didn't add enough planning around

35:03validation. So like go back, edit the

35:05plan, and then let me review again. So

35:08you can add that for any step in Archon.

35:10Or you can have it just decide itself.

35:12Like if you want it to be a little bit

35:13more hands-off, like you want to do

35:15something like the RALF loop, which by

35:17the way, we have a RALF loop workflow

35:19built into Archon for you to use out of

35:21the box,

35:23then you don't have to have the human in

35:25the loop. Like you can have it kind of

35:26like iterate by itself and kind of

35:28critique itself, and then decide when

35:30it's ready to move on. Which obviously

35:32that's a bit more or a lot more

35:34non-deterministic, but you can do

35:36whatever you'd like.

35:37And speaking of that actually, like

35:39workflows in Archon, we have a lot of

35:41Archon workflows that come shipped with

35:44the platform. So when you install

35:46Archon, all of these workflows here are

35:49ready for you to use immediately. So you

35:51can build your own. If you want to

35:53create your own harness, build your own

35:54workflows, you can. But if you just want

35:56to poke around with Archon initially, I

35:57have all of these ready for you guys to

35:59use. So we have like a human in the loop

36:01interactive PRD like walk you through

36:04creating PRDs.

36:06We have the whole like plan to pull

36:08request. We have the RALF loop one if

36:10you want to do the RALF loop in Archon.

36:12It's actually like a beautifully simple

36:13workflow.

36:15Most of it is just prompting here, but

36:16it'll walk you through like creating the

36:17PRD, it'll validate it, and then it'll

36:20go through the whole RALF loop.

36:22And like manage the state through like

36:25JSON and markdown files just like the

36:27RALF loop does, the original RALF loop.

36:30What else do we have? We have the GitHub

36:32issue fix. This is the one that I use

36:34the most out of all of the workflows in

36:36Archon. Because I'll I'll throw this on

36:39the issues that we have in Archon. Like

36:41I'll I'll say like, "All right, Claude

36:43code,

36:44we have a few people that have just

36:45opened up issues for Archon. We got, you

36:48know, like 1082, 80, 76, 72. I want you

36:51to spin up four GitHub issue fix Archon

36:53workflows in parallel, and then monitor

36:56them and let me know when they're done,

36:57and I can review the pull request." Or I

36:59could even have it like run the validate

37:01PR pull request after as well. So yeah,

37:04like I said, I'm always using Archon to

37:06build Archon.

37:10All right.

37:11Visualization is fantastic. I appreciate

37:14it. I assume you're talking about the

37:15the web UI here, but yeah, there's a lot

37:16of work that we've been putting into the

37:18visualization here.

37:21Can I use this with open code? Yeah, so

37:24not yet, but we want to add support for

37:26open code.

37:27So pretty much Archon is going to be

37:29able to be integrated with any coding

37:31agent that supports an SDK.

37:35So like like I said, with Claude, we

37:36have the agent SDK, Codex has their SDK,

37:39Pi has one, open code has one, and Amp

37:42has one. The Gemini CLI does not. That's

37:45another one that I wish I could add, but

37:46they don't have an SDK. Now you can just

37:49run the coding agent in headless mode,

37:52but it's it works better to use the

37:55coding agent programmatically. So this

37:57entire code base is TypeScript.

38:00Almost entirely TypeScript. Because

38:02that's the language most of the SDKs are

38:04written in. So the Claude agent SDK has

38:07both a Python and TypeScript version,

38:10but Codex is only TypeScript. So that's

38:12why I decided to go TypeScript for for

38:14this code base here.

38:16And in the end, like the actual language

38:18you pick doesn't matter a ton, cuz

38:20coding agents can just rock everything,

38:22right?

38:23In fact, TypeScript is even a bit better

38:24for coding agents than Python, because

38:27it has the type safety built right in.

38:31All right.

38:34Cool.

38:36All right. So yeah, there are so many

38:37good questions in the chat here. It's

38:39going to be hard for me to get to

38:40everything, but I'll answer a couple

38:42more here, and then I'll get on to the

38:45demo where I'll I'll install it from

38:46scratch with you guys and show you how I

38:49use it on the day-to-day.

38:51Why have you made this open source?

38:55That's a good question. So open source

38:57has always been a really big part of my

39:00ethos. So as I created started creating

39:02YouTube content in 2024,

39:05I I I really just had a passion for

39:08sharing my knowledge and everything that

39:10I'm diving into with the whole world.

39:12And open source is my way to share it

39:15the most I possibly can.

39:17Because otherwise, if I close source it,

39:20I'm going to tell you guys about it, but

39:22then you won't really care, cuz it's

39:23either going to be some paid product

39:24that you have to shell out a bunch for,

39:26or it's meant for enterprises or

39:28whatever, and then like no one is going

39:29to care about it. So open source is my

39:31way to get the most eyes on it, which it

39:34also is helpful for me, because then I

39:35get the most feedback for it. And

39:38obviously like for my channel, it helps,

39:40because then everyone has a reason to

39:43care about it. If it's not something

39:45that's open source, I can share what I'm

39:46building, but then maybe if you want to

39:48like follow along with my journey,

39:50that's interesting, but otherwise it's

39:51like nothing you can really try

39:52yourself. So, I feel like it just you

39:55either become be kind of like a brick

39:56wall you hit if I make a YouTube video

39:58on Archon, it's like look at this cool

40:00thing I built, and then you can't even

40:01try it. It's like what's the point?

40:03Um so, definitely like with YouTube

40:05being my primary like

40:08platform for sharing things with the

40:10world, like what I share has to be open

40:12source in my opinion. So.

40:15Yeah. It's a fair question, though.

40:19All right. Um yeah, so another person

40:21asked about this and NNN, so I talked

40:23about that already. But yeah, I

40:25appreciate you guys asking about that.

40:28Uh will it include the GitHub CLI or

40:31GitHub Copilot CLI in the future?

40:33So, if they build an SDK, yes, but I

40:35don't think

40:37they have an SDK.

40:40Um oh, wait a second.

40:42Do they actually?

40:44Look at this.

40:46When was this released?

40:48They

40:49Oh.

40:50Okay, so there is a SDK. So, actually we

40:54could integrate GitHub Copilot with

40:56Archon as well.

40:59That's pretty cool. And yep, they

41:00support TypeScript, which is what we

41:01would need for Archon. So, yeah, this is

41:04yet another one. And okay, here's the

41:06really cool thing.

41:07I I haven't really talked about this yet

41:10too much. I mean, this gets more

41:11technical,

41:12but um

41:14let me let me tell you, there is so much

41:16effort that I put into architecting the

41:19initial code base for Archon before I I

41:22ever wrote a single line of code.

41:24So, I have like a sort of like generic

41:27interface implementation for every

41:29single adapter and every single coding

41:32agent.

41:33So, when I built the initial version of

41:35Archon, I supported Claude and I

41:37supported Telegram. So, like Telegram

41:40was a way to talk to it remotely, and

41:42Claude was the only coding agent. But I

41:44built it in a way where it wasn't like

41:45super coupled with that specific tool,

41:47like Telegram or Claude. And so, when I

41:50asked Archon to build support for Slack

41:53as well as Telegram or or GitHub, we

41:56like support you can like talk to Archon

41:57directly in GitHub issues as well. It

41:59just one-shotted it. Like I didn't even

42:01have to iterate, it built it perfectly.

42:03And then when I asked it to build

42:04support for Codex, like following the

42:07pattern that I did for Claude, it

42:09one-shotted that as well.

42:11And so, if you wanted to create a pull

42:12request adding support for GitHub

42:14Copilot, you could. You could literally

42:16one-shot it because the documentation is

42:18there, the coding agent can reference

42:20that if you have like Claude code open

42:22up in the Archon repository or Codex.

42:25And um and then also like it can just

42:27copy how we already have it set up for

42:29an existing coding agent. So, we have

42:31Archon set up in a way where the code

42:33base is very easy to evolve because

42:36everything is set up as like super easy

42:38patterns to understand. Like the coding

42:40agent can understand very easily, like

42:42here's how I add a new coding agent,

42:44here's how I add a new adapter, uh

42:46whatever we want to do to evolve the

42:48system.

42:50Yeah.

42:53Um can workflows be told to obey token

42:56rates and limits?

42:58So, I think that's actually something we

43:00have an open issue for cuz there's not a

43:03way to like set a max budget for tokens,

43:06but it would be pretty easy to add that.

43:08Like on a per node level, it's like I

43:10want you to stop if this is taking more

43:12than 100,000 tokens cuz this that

43:14definitely means that like the coding

43:15agent is going off the rails here or

43:17whatever. So, we don't have support for

43:19that yet, but um I think I mean, I don't

43:22I don't really want to like go on a hunt

43:23here and try to find this exact one,

43:26but uh there are there is definitely an

43:28issue that we have out for for that.

43:31Let's see if I can find it. So, mo- most

43:34of these issues that we have right now

43:35are actually created by uh Rasmus. So,

43:37like I said, he's one of the guys that's

43:39been helping me an insane amount with

43:40Archon. So, most of these things are

43:42like just us kind of listing out things

43:44that we want to

43:46uh improve in in Archon.

43:50Um we had to port over so, we were

43:52working on a kind of like a private repo

43:53that was like just for the Dynamis

43:55community at one point. So, we had to

43:57port over a lot of the issues. I don't

43:59know. We might have changed the names

44:01for some of them,

44:03but that's definitely something we have

44:04on our radar.

44:06Um yeah, not sure if we I can't find it

44:10exactly right now. But yes, that's one

44:12of the things that's just like all of

44:13the different parameters that we have in

44:14like Codex and Claude code, we want to

44:16make sure that we can support those so

44:18you can build it directly into the

44:19configuration for each node of an Archon

44:22workflow.

44:25All right. Yeah, and so yeah, one thing

44:27I I want to mention really quick is like

44:30when we were first working on Archon,

44:33we were doing it like as an internal

44:35project in the Dynamis community. So, we

44:38called it the actually I still need I

44:40still need to archive this repository

44:42here. But we have the remote coding

44:43agent. It's funny cuz Archon the new

44:46version of Archon started as a a

44:50resource that I built for the Dynamis

44:52agentic coding course because it it

44:54originally wasn't a harness builder. It

44:56was just a platform that allowed you to

44:58talk to Claude code or Codex

45:01uh in a remote environment, like Slack,

45:03GitHub, or Telegram, like I was talking

45:05about earlier.

45:06So, that's how it started, and then it

45:08evolved into this beautiful thing where

45:11now you can like build any AI coding

45:13workflow as an Archon harness.

45:15Uh but this this is like the origin of

45:17Archon. So, it's cool. For everyone in

45:19the Dynamis community, you guys got to

45:21like see it evolve to the point where it

45:23is now, and obviously you got early

45:25access to it. Another thing is um within

45:28the Dynamis community, I am going to be

45:31doing a lot more workshops the next

45:33couple of months, like getting like

45:34really deep into using Archon, like

45:36building custom workflows. I'll show you

45:38guys more like how to build your own

45:39harnesses. Um it's it's always the place

45:42to be if you want to like be a part of

45:43Archon and get the the inside scoop on

45:47how to use the tool the best and like

45:49the evolutions that I'm doing to it. Uh

45:51so, I yeah, I just want to call out

45:52really quick for Dynamis. I'm going to

45:54put a link in the chat here. Uh I

45:56actually have a little live stream

45:58special for you guys

46:00for Dynamis. So, it's a 10% off the uh

46:04public price for the community. This

46:06discount here is going to go away

46:10uh by the time our live stream is done

46:12here. So, it's a special for literally

46:13just this live stream. So, if you like

46:16what I'm working on with Archon, if you

46:18want to be a part of of this, and um

46:21also all the other course content that I

46:23have in Dynamis, definitely check this

46:25out. I put the link in the chat just

46:26now. So, another big thing that I did in

46:28Dynamis recently is I did a 4-hour boot

46:32camp on building your own AI second

46:34brain. So, I took the entire second

46:36brain that I built for myself that

46:38literally saves me 20 hours a week, and

46:40I showed you how to build it from

46:41scratch. So, I did a a live workshop,

46:43and I'm also turning this into the third

46:46course for Dynamis as well. We also have

46:49the AI agent mastery and agentic coding

46:52courses, and I referenced the agentic

46:54coding course already because that was

46:55the origin of the this new version of

46:57Archon.

46:58So, yeah, lot a lot of value packed into

47:01Dynamis. I do weekly workshops as well.

47:03And so, that's where I'll be doing some

47:05more stuff with Archon if you are

47:07interested.

47:08So, all right. Uh with that here, I'm

47:11going to go ahead and show you guys how

47:13to install Archon from scratch.

47:17And uh like I said, I'm going to be

47:18doing it on a VPS

47:21just because I don't want to wipe my

47:24installation on my computer again. I I

47:27already reinstalled Archon from scratch

47:29um four times this week as I was testing

47:32things to get ready for the open source

47:34release. And so, I'm going to install it

47:37on the VPS where I'm going to build my

47:39dark factory. This is the exciting like

47:42last part of the live stream that I'll

47:43be covering with you guys here. Um yeah,

47:45I'm I'm actually pretty stoked for this.

47:47I don't know if you guys know what a

47:48dark factory is. Um it's

47:51kind of like a a term that was coined in

47:53the last Well, actually it was a a term

47:56that was coined in the late uh 1900s.

47:59Uh that sounds weird. It like 1990s, I

48:02think, like the idea of a dark factory

48:05is um you have a factory without lights

48:07because it's robots running the entire

48:08thing. And then recently people have

48:10been talking about like dark factory as

48:12it relates to code bases. And basically,

48:15a dark factory is a code base that

48:17self-evolves. Like AI is the only one

48:19writing code ever on the code base. And

48:22so, an experiment that I want to run,

48:25I'm giving you a little bit of a teaser

48:26to what I'll talk about at the end of

48:27the workshop here, an experiment I want

48:29to run is a public dark factory code

48:33base where every single evolution of the

48:35code base, like every pull request,

48:37every release, is managed by Archon

48:40workflows.

48:41Because we can build our process like we

48:44can define exactly how we want to manage

48:46issues and pull requests and releases,

48:48we can define it as Archon workflow. So,

48:50we basically use Archon as a dark

48:52factory harness, which sounds kind of

48:56it sounds silly. Uh but I'm actually

48:58like really excited to try this as a um

49:00kind of like public experiment here.

49:02Like I I literally wanted to get to the

49:04point where like anyone can create an

49:06issue, and then I'll have Archon like

49:08figure out is this an issue that we

49:09should uh address for this code base,

49:11and then it'll handle it automatically

49:13all the way to like pull request and

49:15reviewing and merging into the main

49:16branch.

49:18So, we'll talk about that, but obviously

49:20we got to get Archon uh spun up for the

49:22first time here. So, I'll show you guys

49:24the installation. I'll show you what it

49:26looks like to run workflows with Archon.

49:28And then I want to build a workflow with

49:31you guys. And then we'll do some of the

49:33dark fa- factory fancy stuff at the end.

49:37All right. So, I'm going to go into the

49:40readme here.

49:42So, I'll put a link to Arkon in the chat

49:45again.

49:47If you guys want to follow along and

49:48install with me on your machine or a

49:50VPS.

49:51If we scroll down to the get started,

49:54this is where we're going to work right

49:55now.

49:56>> [snorts]

49:56>> So, literally all you have to have

49:58installed before you run Arkon is your

50:02coding agent and the GitHub CLI. In

50:04fact, I list bun as a prerequisite, but

50:06the setup process is going to even

50:08install this automatically

50:10if you don't have it yet.

50:13So, I do list Claude code as a

50:14pre-wreck. You can use Codex as well.

50:17That's also supported with Arkon. It's

50:19just not quite as stable right now as

50:21Claude code. So, I'll definitely be

50:23updating the docs as I address that or

50:25as we address that.

50:27But then obviously the GitHub CLI is an

50:29important dependency as well.

50:32You don't need the GitHub CLI for every

50:34Arkon workflow.

50:36But um yeah, being able to work with

50:39GitHub is a very core part of Arkon

50:42because most workflows are relying on

50:45having the issue or pull request be like

50:48the starting point, right? Like

50:49reviewing pull requests or fixing GitHub

50:51issues. Most of the time when I'm

50:53working with Arkon, it's it's dealing

50:55with artifacts in GitHub essentially,

50:57right? So, that's why I have that as a

50:59prerequisite.

51:01So, once you have those things

51:02installed,

51:03>> [snorts]

51:03>> um we have the binary. So, we're working

51:05on making it so there's like a single

51:07install script. So, you can do this if

51:09you want. I'm just going to show you

51:11that my my approach of like cloning the

51:13repository and setting up from there cuz

51:14that's what I'm more comfortable with.

51:16This is like a newer thing that we're

51:17working on to make it like even easier

51:19to set up Arkon. So, I'm going to

51:21install it this way. So, I'll I'll copy

51:23this command to clone the repository and

51:25I'll do that within my VPS here.

51:30So, there we go. Clone Arkon and then

51:31obviously change my directory into Arkon

51:34and then check this out. Um so, I'm

51:35going to run in dangerously skip

51:37permission mode just because I don't

51:39want to have to deal with approving

51:41things right now.

51:42Open up the folder. Uh yes, I accept.

51:45There we go. All right. Now, watch this.

51:48All I have to do is say set up Arkon.

51:51I don't have to say anything else. It's

51:53going to walk me through the entire

51:54process. So, you if you have the

51:55pre-reqs and then you clone the repo,

51:57you just go into your Claude or Codex

51:59and you say set up Arkon and that's it.

52:02So, first it's going to uh see like what

52:05we already have set up. It's going to

52:06realize that we haven't set up anything

52:09and then what it should do is it should

52:10load the Arkon skill. So, we have this

52:13skill that like walks it through how to

52:15help us with Claude or with the the

52:18whole setup of Arkon with Claude.

52:23All right.

52:24>> [cough and clears throat]

52:26>> I guess I have to wait for it to to work

52:28here.

52:30And by the way,

52:32if we see it um I'm just did like

52:34control O to see the output here.

52:36It's just like understanding the code

52:38base. I'm actually surprised it didn't

52:39load the Arkon skill.

52:42Um you're in the let's see. What is it

52:46saying here? No, all right. Hold on.

52:49Load the Arkon skill and walk through

52:52the setup. I've actually never had to do

52:53this before,

52:55but it uh for some reason didn't load

52:56the Arkon skill. So, I'm just going to

52:58tell it to do that.

52:59But we're still going to get the same

53:00effect here. What will be super easy to

53:02go through everything.

53:05All right.

53:07There we go. So, now it's reading the

53:08setup guide. So, this is like a part of

53:09the skill

53:10that tells it how to walk us through

53:12everything.

53:15So, it's going to check all the

53:16prerequisites like making sure we have

53:17get and bun, things like that, which we

53:19already do on this machine.

53:22Um there we go. So, now the first

53:24question that it'll ask you in the setup

53:26is where like what's the repository that

53:28you want to use to work on with Arkon,

53:32right? So, like when we use Arkon, we

53:34want it to invoke workflows on another

53:36repository cuz we're going to use it to

53:38work on something else. So, that's why

53:40it says this should be your own project,

53:42not the Arkon repo.

53:44And so, I can either

53:46clone a repository from GitHub or I can

53:49just give it a local path if I wanted to

53:51work on a a project that I already have

53:53installed on my machine. So, number two

53:55is actually probably what you would end

53:57up doing cuz usually there's already

53:59something you're in the middle of

54:00working on that you want to use Arkon to

54:02like build a harness around, right? Um

54:05now, since I'm on a VPS, I am just going

54:06to clone something from GitHub here.

54:09So,

54:10I will select that and then it just says

54:12like please provide the repo URL.

54:14Um I need to actually find one here.

54:17I'll I'll just do let me open this up.

54:20I'll pull some like random

54:23um

54:24let's see. Let's just do this one. I'll

54:26just copy this repo. I just need kind of

54:27like a random one here to register. So,

54:30what this is going to do is it's going

54:31to register your first repository with

54:34Arkon.

54:35So, whenever you

54:37um run the Arkon CLI on a repo for the

54:41first time, it will register. So, like

54:43Arkon has a database under the hood or

54:46behind the scenes that like keeps track

54:48of the projects that you're using Arkon

54:49with and like all the conversations

54:51you've had where you've invoked Arkon

54:53workflows on those code bases. So, you

54:55do the automatic registration that way

54:57and then also in the web UI like I

54:59showed earlier, you can register

55:01projects this way, too. So, you just

55:03click the plus icon here and then you

55:05can give it the GitHub URL or the local

55:06path. So, also in the web UI, you can

55:08register your repos.

55:10And the cool thing is for all the

55:12projects that you have registered in the

55:14web UI, the agent automatically

55:16understands. So, if I say like what

55:17projects and workflows do I have,

55:21the Arkon agent that runs in the web UI

55:24either using Claude code or Codex, it it

55:26has the context injected for all of the

55:30Arkon workflows it can invoke and the

55:32different projects that it can invoke it

55:33in. And so, when we say like, "Hey, I

55:35want to fix this GitHub issue for this

55:37project." it will know like, "Okay, let

55:39me use the fix GitHub issue workflow and

55:42I'll route it to the rag YouTube chat

55:45repository.

55:47And the the web UI is one of the things

55:49that we can spin up once we go through

55:50the setup here. So, first it asks what

55:52platforms do you want to set up. So, the

55:53CLI is always included by default. I can

55:56also set up you know, like Telegram. So,

55:58you just do you know, like

56:00um enter to select or unselect. Um for

56:03this, I'm going to keep it really simple

56:05and just do the CLI.

56:07Um yeah, but you can set up like all the

56:10adapters that we support in Arkon right

56:12here. So, you specify the ones that you

56:14want and then in a little bit it'll go

56:16through a process where we can give our

56:17API keys. But it's a separate process

56:20because we don't want to just send our

56:21API keys directly into a coding agent or

56:24an an LLM. So, I'll submit these answers

56:26here.

56:28So, I just want to set up the CLI so I

56:29can run workflows there. So, the CLI and

56:32the web UI obviously don't require you

56:34to have any additional parameters or API

56:37keys. So, they'll just come they'll work

56:38right out the gate.

56:40So, then what it does here is it creates

56:43Arkon as a global command. So, that way

56:45we can just run you know, like Arkon the

56:47CLI to invoke workflows. But trust me,

56:49you're never going to do this yourself.

56:50You're just going to have your coding

56:52agent run Arkon workflows. So, you don't

56:54really have to understand the CLI

56:56yourself. That's the beauty of having an

56:58Arkon skill is your coding agent can

57:00load the skill and then it knows how to

57:02invoke the Arkon CLI to kick off

57:04workflows for you.

57:10All right. And so, now what happens is

57:13it helps you configure your credentials.

57:16But it runs this in a separate terminal

57:18because we don't want our coding agent

57:20to see our API keys. That would be a

57:22huge security risk.

57:24And so, it it will automatically spin up

57:27a new terminal for you to go through the

57:29setup. But depending on your operating

57:32system, it might not there might not be

57:34support to like automatically start a

57:35new terminal.

57:37And so, you might need to just use do

57:40this yourself. And so, for me because

57:42I'm running in a VPS, there's no like

57:44automatic terminal spin up. So, I just

57:46have to connect into the machine again.

57:48So,

57:50let me do that. So, I got to remember

57:52the path to my SSH key.

57:55Uh

57:56there we go.

57:59Goodness, I can't type today.

58:01Gosh. All right. And then

58:03um I forgot the IP of my machine. Shoot.

58:06What is the IP address of my machine?

58:08Here, I'm going to pull this up on

58:10another monitor here.

58:13Um

58:17All right. Pull this up.

58:23There we go. Okay.

58:26All right. So, SSH in and then I will go

58:28to

58:31the path here, dark factory Arkon.

58:36Wait, that's not it.

58:41Dark factory

58:43uh oh yeah, capital A. Okay.

58:45All right. So, then all you have to do

58:47in the separate terminal is run the

58:48Arkon setup command and it it walks you

58:51through this here. If it doesn't spin up

58:53the terminal automatically, you just do

58:54that yourself. And now we just go

58:56through the setup process. So, I'll zoom

58:57in on this here. So, first what database

58:59do you want to use? SQLite is the

59:01easiest to set up and so, I recommend

59:04that. You can use Postgres though if you

59:06want to have an external database you

59:08connect to. Uh which shoot, I

59:10accidentally entered that. So, let me go

59:12through the setup again. So, SQLite and

59:14then here's where it asks what coding

59:15agent you want to use. And so, these are

59:17the two we support right now. Like I

59:19talked about, we want to add more as

59:20well.

59:21So, space to select and then enter to

59:23confirm. And then it asks, "How do you

59:25want to authenticate with Claude?"

59:27And so, there's three options here. We

59:28can give an OAuth token. Like if you

59:31want to just run you know, the Claude

59:32like setup token command to get an OS

59:35token you can give your API key which I

59:36wouldn't recommend cuz it's going to get

59:37expensive and then like I said we are

59:40allowed to use our anthropic

59:41subscription with Claude code with the

59:43Claude Asian SDK and Archon. So I'm just

59:45going to use my global off. So on this

59:47instance I've already authenticated with

59:50Claude.

59:51If you use Claude code a lot and you're

59:53installing Archon on your machine then

59:55the authentication is already set right

59:56like you don't even have to set up any

59:58other environment variable.

1:00:00And then it asks what platforms I want

1:00:02to connect as in what ones do I want to

1:00:03enter in an API key for. I'm not

1:00:06actually going to I mean I guess I'll

1:00:08just do GitHub right now just so I can

1:00:10show you one of these. So you do

1:00:13space to select and then enter to

1:00:16confirm. And then what it'll do is for

1:00:18each of the platforms that you select

1:00:20that you want to install Archon with.

1:00:21This is very similar to the open claw

1:00:24setup.

1:00:25If you guys have installed open claw

1:00:27before is for each one of the platforms

1:00:29it'll like give you instructions. Like

1:00:30here's how you get your your GitHub

1:00:32personal access token.

1:00:34And so I'm going to off camera obviously

1:00:38go ahead and get that and copy it.

1:00:40So let me go to

1:00:43there we go.

1:00:45So I'll paste it in because it

1:00:46automatically hides it.

1:00:49Which is good wouldn't want to show that

1:00:50on the live stream

1:00:51but yeah gives you instructions for how

1:00:53to do that. It'll be the same for every

1:00:54single one of your platforms that you

1:00:56configure. And then another thing

1:00:58obviously for the sake of security is we

1:01:00for each of the platforms that we set up

1:01:02like GitHub and slack and telegram we

1:01:05want to have a comma separated list of

1:01:07users that are allowed to invoke Archon.

1:01:10Because if I have Archon running on a

1:01:11public repository I don't necessarily

1:01:14want anyone to be able to just say like

1:01:16you know at Archon fix this issue

1:01:18because then it's spending my tokens and

1:01:20it's not me invoking it. So I'll just

1:01:22say like this is the only user that's

1:01:24allowed to use it.

1:01:25And then you can change the mention

1:01:27names like in GitHub you do at Archon

1:01:29and then you give it the request like

1:01:31that's how you talk to it. So I'm good

1:01:32with with that as the default.

1:01:35So I don't need to set up anything

1:01:36separate.

1:01:38And then the other thing is it also for

1:01:41that repository that we we registered.

1:01:44So this repo that I found just as like a

1:01:47a random one where was it? Yeah this

1:01:49one. So like I I gave it this to

1:01:51register as the first repository.

1:01:54We can also copy the Archon skill into

1:01:57that project.

1:01:58And the the reason I want to do that is

1:02:02then I can open up my coding agent

1:02:04directly in that code base and since the

1:02:07Archon skill is there it knows how to

1:02:10use the Archon CLI. So that way I don't

1:02:12have to open my coding agent in the

1:02:14Archon repo in order to work on another

1:02:17repository.

1:02:18Because as long as the Archon skill is

1:02:20there the Archon CLI is a global CLI. So

1:02:24we can invoke it from anywhere on our

1:02:26machine. Like I can just open up my

1:02:28terminal right here and just say like

1:02:29Archon right and then that automatically

1:02:32works

1:02:33because it's a globally registered

1:02:34command but our coding agent only knows

1:02:36how to use it like use all the different

1:02:38commands and options if we have the

1:02:40skill. So I just want to copy it over

1:02:42into the code base. So

1:02:45All right and then the docs directory

1:02:47not to worry about that right now. And

1:02:48there we go our setup is complete. And

1:02:50so we go through all of this

1:02:52and then we go back here

1:02:55and we just say done. So like we

1:02:56finished and now it's going to validate

1:02:59all of the credentials

1:03:01and make sure that we're good to go and

1:03:03then it'll actually test the Archon CLI.

1:03:05So it'll run a workflow for us to make

1:03:08sure that everything is all configured

1:03:11properly.

1:03:12So we'll let that run here as well.

1:03:19All right.

1:03:21So here we go running a quick test here.

1:03:22So we can see that it is using the

1:03:24Archon CLI. So it thanks to the skill it

1:03:27knows how to. You never have to worry

1:03:29about running it yourself. You just ask

1:03:30the coding agent to do so. So it's

1:03:32running the Archon assist workflow

1:03:34and it is doing it on our repo that we

1:03:38registered the first repo that we

1:03:39registered.

1:03:41And so it's just you know say hello like

1:03:43a really basic test to make sure that it

1:03:44works. And there we go it's good.

1:03:46So yeah I'll copy the skill over. I

1:03:48guess it asks us here as well. Maybe

1:03:50that's something to touch up but anyway

1:03:52so every single time that you

1:03:56invoke a workflow from the CLI it's

1:03:59going to run as a background process. So

1:04:01Claude code or code X like it has access

1:04:03to

1:04:05the full logs from the workflow. So you

1:04:08can like ask it for a status if it's a

1:04:10longer running workflow or you can say

1:04:12like hey summarize what happened in the

1:04:14workflow like it's

1:04:15able to basically you know communicate

1:04:17with the workflow that it runs because

1:04:19all the logs are right there in the

1:04:21background process.

1:04:23So there we go our setup is complete. So

1:04:25it tells us what is configured and then

1:04:27it gives us next steps as well. So take

1:04:29a look at this.

1:04:30We are immediately ready

1:04:32to use Archon in our code base now and

1:04:36we can register it with any other code

1:04:38base we want as well. So like I can just

1:04:40say like you know

1:04:42I'll go into my speech to text tool here

1:04:43and I'll say like use the GitHub issue

1:04:45fix workflow to fix issue number two on

1:04:49XYZ repository right? Like I can send

1:04:51off this request

1:04:54and it will automatically if I you know

1:04:56give it obviously the path to the repo

1:04:58here it'll use the Archon CLI to run the

1:05:02GitHub fix issue workflow on this repo

1:05:05and then that also automatically

1:05:07registers that repository with Archon so

1:05:09it knows about it going forward if the

1:05:11repository wasn't registered already.

1:05:14And then the other thing is wherever I

1:05:16have the Archon skill copied which if

1:05:19you want to copy it yourself in the

1:05:21Archon repo it's just within the dot

1:05:23Claude skills. So this and I'll put a

1:05:26link to this in the chat right now as

1:05:28well. This skill

1:05:30which you can just ask your coding agent

1:05:32to copy it into your new like whatever

1:05:35repo you want to start using Archon with

1:05:37but this skill tells it how to use the

1:05:39CLI. Like this is literally the only

1:05:41requirement. So another thing is if you

1:05:43want to use your second brain with

1:05:47Archon all you have to do is put this

1:05:50skill into your second brain repo.

1:05:53And then it'll immediately be able to

1:05:54start using Archon workflows. So you can

1:05:56basically add Archon as the coding arm

1:05:59for your second brain. So you can create

1:06:01your own harnesses your own workflows

1:06:03and then you can tell your second brain

1:06:04to invoke it on whatever repo you you

1:06:06want as long as you give it the path and

1:06:08it loads the Archon skill then it knows

1:06:09how to use the CLI.

1:06:12And so for anything here if it's like

1:06:14confusing like okay what repo do I open

1:06:15up or like how do I get the skill or how

1:06:17do I register projects like you

1:06:19literally just ask Archon and it knows

1:06:22how to do everything because the skill

1:06:23walks it through everything. So let me

1:06:25actually show you this here. So in this

1:06:27other terminal here

1:06:29I'm going to

1:06:31um

1:06:32clear and then so it says I can change

1:06:34my directory into the Claude memory

1:06:36compiler.

1:06:38And then I can run Claude to launch

1:06:40Claude here. So I'll do

1:06:43the dangerously skip permissions again.

1:06:46Or I guess it's not here. I'll just have

1:06:47to type it out manually. So Claude

1:06:49dangerously skip permissions.

1:06:53And then you can see that since I copied

1:06:55over the skill to this repo during the

1:06:56setup I can just say load the Archon

1:06:58skill. Not that you have to say this

1:07:00explicitly every time but I'm just

1:07:01demonstrating that like we have the

1:07:03Archon skill in this repo. So now it's

1:07:05going to know how to invoke any

1:07:07workflow.

1:07:09And then within the repo here okay we

1:07:12actually have a couple of issues. Cool.

1:07:15This 6 p.m. thing is very edgy. I've

1:07:18never actually seen this issue before by

1:07:20the way cuz this is a public repo.

1:07:22Um

1:07:25That's okay. Massive token consumption.

1:07:28Interesting. [snorts]

1:07:30All right well I'm trying to find one

1:07:31that's actually like

1:07:35Um see it exists. Okay well let's try

1:07:37this one. This is kind of random but I'm

1:07:39just going to pick a random issue like

1:07:40number one here. So like watch this. All

1:07:43I have to do is say

1:07:45I want you to use Archon to fix issue

1:07:48number one.

1:07:49That's it. It's so simple cuz it knows

1:07:52the workflows it has access to. It's

1:07:54going to pick the right one right? Like

1:07:55it's going to pick the Archon fix GitHub

1:07:57issue. It knows how to use the CLI.

1:07:59Boom. That is it and we have this full

1:08:02process running now. So Archon fix

1:08:05GitHub issue is running in the

1:08:06background. It's going to check on the

1:08:08progress periodically and

1:08:11um

1:08:12Oh hold on.

1:08:13The workflow failed because you're not

1:08:15logged into the GitHub CLI.

1:08:17Oh that's a bummer. Okay hold on. I

1:08:19guess I have to do that. I thought I

1:08:20already did the login here.

1:08:23Um

1:08:25Let's see. Paste authentication token.

1:08:27Hold on. I'm going to do this off camera

1:08:29quick.

1:08:31Um I thought I already did that part of

1:08:33the setup so it's kind of weird that it

1:08:35says but I'm I'm just going to try the

1:08:36login here.

1:08:39All right. So

1:08:42Yeah see it says I'm already logged into

1:08:44this account. I think Claude might be

1:08:46tripping right now.

1:08:48Let me try resuming the conversation.

1:08:51Let's see.

1:08:53I logged in. Also it says I already was

1:08:57so I am confused. There might be

1:08:59something else I forgot to or I messed

1:09:00up in my configuration on this machine.

1:09:03Let's see.

1:09:06But it's cool like it can just rerun the

1:09:08workflow right? Like we can just talk to

1:09:09Claude as we normally would.

1:09:12And so it can use Archon as a tool just

1:09:14like it would use sub agents or just

1:09:15like it would use skills.

1:09:17It says it exited again.

1:09:20The error is actually from the Claude

1:09:22code off. The not logged in is coming

1:09:23from the Claude code agent that Archon

1:09:25spawns.

1:09:27Um

1:09:32Check Claude.

1:09:34Oh, I think I know what it is. I think

1:09:36it's because of the

1:09:39Uh it's because of this specific repo. I

1:09:41have the dot Claude folder with the

1:09:43settings.json.

1:09:45If I just remove this

1:09:48I might just do the demo on my computer

1:09:50instead of this VPS cuz I I think it's

1:09:53just this specific code base that I have

1:09:55some of my Claude code hooks that are

1:09:56running. It's a

1:09:57whole thing.

1:09:59Oh, wait. Oh, wait. I know what's wrong.

1:10:01It's because I have this already set up

1:10:03to use MiniMax M2.7 for the dark factory

1:10:06stuff I was going to show you guys.

1:10:07That's my bad.

1:10:09Uh okay, I have to I have to demo this

1:10:11on another machine here. And by the way,

1:10:13we were using MiniMax 2.7 for

1:10:15everything.

1:10:16>> [laughter]

1:10:17>> Um cuz I was getting some stuff set up

1:10:18ahead of time for the live stream here.

1:10:21Sorry guys, I'm complicating things more

1:10:23than I

1:10:24um

1:10:26I then I need to because I I'm I have

1:10:28some other things prepared for you guys

1:10:30here. So okay, here's what I'm going to

1:10:32do. I'm going to show you a

1:10:33demonstration by doing it right from

1:10:35Arkon. So um this will be better anyway

1:10:38cuz then I can show you guys the web UI.

1:10:40Uh which by the way, if you want to

1:10:41start the Arkon web UI, all you have to

1:10:44do is go into the Arkon code base and

1:10:46say start the back end and front end of

1:10:48Arkon. That's it.

1:10:51Um start the back end and front end of

1:10:52Arkon. I guess my speech text got cut

1:10:54off, but that's all you have to do. So I

1:10:57hope that like the pattern is clear here

1:10:59for your setup, for getting things up

1:11:01and running, for running Arkon. It just

1:11:03comes down to like the Arkon skill

1:11:06guides it through everything. All

1:11:08workflow execution managing the

1:11:10application, super super easy.

1:11:13So within my Arkon here let's go ahead

1:11:16and uh zoom in a bit. Let's say I just

1:11:18wanted to handle an issue. So this is

1:11:19going to be kind of meta cuz I'm using

1:11:21Arkon to improve Arkon. Um but like I

1:11:23said, you could do this on any code base

1:11:25where we have the Arkon skill.

1:11:27Um so I'm going to find an issue.

1:11:29There's quite a few issues that have

1:11:30been created in the last couple of days

1:11:31cuz we're we have a lot of eyes on Arkon

1:11:33right now. We're at 16.2 thousand stars,

1:11:36which I'm honored by the way. Um

1:11:40Okay.

1:11:41Uh let's see. So okay, this is a good

1:11:43one. So chat UI fails silently when the

1:11:45Claude OAuth refresh token is expired.

1:11:48So this is issue number 1076.

1:11:50So watch this. I'm going to go in this

1:11:52just like I was trying to do in the VPS.

1:11:55I just have things uh configured. We'll

1:11:57talk about the dark factory stuff in a

1:11:59bit cuz I want to use MiniMax um

1:12:02anyway. So I'll go in here and I'll say

1:12:04uh use Arkon to fix uh issue number

1:12:071076.

1:12:10There we go.

1:12:12All right. Now now we'll see it in

1:12:13action. So uh again, it'll load the

1:12:15Arkon skill. Well, I guess first it'll

1:12:18view the GitHub issue. And then yep,

1:12:20load the Arkon skill so it knows how to

1:12:22invoke workflows and then it'll kick off

1:12:24the workflow.

1:12:25So Arkon fix GitHub issue and then uh

1:12:28we're doing it in a branch. So it's

1:12:29going to do this in a work tree so we

1:12:31have isolation. And there we go. So now

1:12:33the workflow is running in a background

1:12:35process. So Claude code has support for

1:12:37this. I don't know if Codex does as well

1:12:39cuz I haven't used Codex in a while if

1:12:41I'm going to be honest, but if you click

1:12:43into the shell here like if I if I press

1:12:46the down arrow and then hit enter, I can

1:12:48see the details and the logs of the

1:12:51workflow as it's running. So we can

1:12:54monitor it here.

1:12:56And then the cool thing is because it

1:12:57runs as a background process, I can

1:12:59continue. I can run more Arkon

1:13:01workflows. I can keep just talking to

1:13:03the agent here. I can also say give me a

1:13:05status update. So it can look into the

1:13:07logs for the background process and then

1:13:09tell me what stage of the workflow it's

1:13:11in. So if I want to check in cuz this is

1:13:12a longer running workflow, it can say

1:13:14like oh, it's currently investigating or

1:13:16it's in the middle of classifying the

1:13:17issue. So it it reads like I know it

1:13:19looks kind of long here, but this is the

1:13:21the logs that are stored internally on

1:13:23my machine for the background process.

1:13:25And then it says, all right, cool. So

1:13:27work tree is created. It uh extracted

1:13:29the issue number and it classified the

1:13:31issue. So it's currently running the web

1:13:33research step. And if we go to the web

1:13:36UI here

1:13:37I can actually see that. So take a look

1:13:39at this. We have it currently running so

1:13:41I can see it in my chat. If I go to the

1:13:43mission control right here, I can also

1:13:45see all of my running UI but especially

1:13:48as you're building your own custom

1:13:50workflows, it can be really really

1:13:52useful when you're first debugging

1:13:54things to dive into a workflow log and

1:13:57like make sure that things are actually

1:13:58happening as you intend them to.

1:14:27And so we can see the logs. I know like

1:14:29this one specifically the the web

1:14:31searching it it's um a longer step, but

1:14:34like we can see all these tool calls

1:14:36come in live

1:14:37as uh as it's working.

1:14:40Um and then we can invoke a ton of

1:14:42different workflows in parallel and we

1:14:44can watch them all here. And we can

1:14:45click between the logs for them. Uh it's

1:14:47pretty cool. So yeah, and so we can

1:14:51monitor it here or in the web UI.

1:14:54And uh you know what? Just for the sake

1:14:56of getting kind of fancy here

1:14:58I'm going to uh I'm going to show you

1:15:01guys how I how I actually use Arkon

1:15:03every single day. Cuz here's the thing.

1:15:06I am not just working on one issue at a

1:15:08time with Arkon.

1:15:10Um the fix GitHub issue workflow is my

1:15:13most often used, but I'm using it in

1:15:15parallel a lot. So take a look at this.

1:15:18I'm going to go into my speech-to-text

1:15:19tool.

1:15:20I'm going to say

1:15:22Okay, I also want to handle more issues.

1:15:26So let's see here.

1:15:27Uh let's tackle issue number 1067.

1:15:32Um

1:15:34Let's see. Let's also do

1:15:38um 1082

1:15:42uh and 1087.

1:15:45All right. So I'm going to send this in.

1:15:47Uh by the way, usually I'll do even more

1:15:49than this at once, but I just want to

1:15:51make sure that I don't hit my rate

1:15:52limits for Claude right now.

1:15:54Cuz it like I mean, we're doing a lot of

1:15:56work in parallel here. Not that Arkon is

1:15:57token inefficient. In fact, we've been

1:15:59doing a lot of things to make it more

1:16:01token efficient, but still like there is

1:16:03a a bunch that I'm doing at the same

1:16:05time right here.

1:16:07And watch this. Not only can I say, you

1:16:09know, spin up the workflows in parallel,

1:16:11but I can say I want you to, you know,

1:16:14run these workflows in parallel. I want

1:16:16you to wait until all of them are done.

1:16:18So monitor the workflows until we have

1:16:20pull requests for every single one of

1:16:22them. And then once we have pull

1:16:23requests for every single one of them,

1:16:25then I also want you to run the validate

1:16:27PR workflow on all these in parallel.

1:16:30And then when it produces the comments

1:16:32in the PRs, I want you to view the pull

1:16:34requests, look at all of the issues that

1:16:36we need to address. I want you to

1:16:37address them and push the changes to the

1:16:39branches for the pull request.

1:16:41And I could even combine all this into a

1:16:43single workflow if I wanted as well. But

1:16:44it's like look at how comprehensive this

1:16:46is. We are going all the way from issue

1:16:48to a final validated pull request.

1:16:52And we're doing it in parallel. So it's

1:16:53spinning up all of these at the exact

1:16:55same time

1:16:57as background processes. So now if I go

1:16:58into the four shells that I have open

1:17:00here in Claude code, we can see that

1:17:02they're they're all currently running.

1:17:03And then I can go hit enter to view the

1:17:05logs for any one of them.

1:17:07And um then of course, I can see them

1:17:09all running in the web UI as well. So

1:17:11let's go back to the dashboard and there

1:17:13we go. We got all four of them running.

1:17:15So we have each of them running in the

1:17:16last 20 seconds here and then this one

1:17:18that's been going on for 5 minutes now.

1:17:21Pretty cool.

1:17:24So yeah, we're still doing the web

1:17:25research on this one here.

1:17:28But yeah, that that's how I use Arkon on

1:17:30the day-to-day. Like most of the time

1:17:32when I'm working on a code base

1:17:35I'm I'm going to be filing things as

1:17:37issues, whether it's a bug that I'm

1:17:39going to be working on or it's a new

1:17:41feature that I want to add. Like both of

1:17:43those those fit as GitHub issues.

1:17:46And it's also nice because this is kind

1:17:47of like my personal mission control for

1:17:49everything I want to work on. And so

1:17:51that's why we see like Rasmus opening up

1:17:53so many GitHub issues. Like we're using

1:17:55this more than you know, other people

1:17:57are because this is where we document

1:18:00all the things that we want to work on.

1:18:01Like Rasmus is opening up issue after

1:18:03issue

1:18:04for you know, bugs and feature requests.

1:18:06And so it's also really nice because

1:18:09GitHub is where we're going to keep

1:18:10track of the versioning of Arkon. Like

1:18:13as you evolve your code base and you

1:18:14want your coding agent to remember

1:18:16things that you've worked on in the

1:18:17past, you're going to rely on Git

1:18:19commits. I actually love using the Git

1:18:21log as long-term memory for my coding

1:18:23agents. And so I'm already keeping track

1:18:25of all my work in in uh as the Git

1:18:28commits. And so I might as well like

1:18:29track things in issue in issues as well

1:18:31for like the upcoming work that I have.

1:18:33So you can of course hook in like a MCP

1:18:35server or skill for Arkon to use another

1:18:38external platform like Linear or Jira

1:18:40for your task management. I personally

1:18:42just love using uh GitHub as my task

1:18:45management. So that's why the GitHub CLI

1:18:47is so crucial for me for pretty much all

1:18:50of my Arkon workflows.

1:18:52Now of course, there are Arkon workflows

1:18:54that don't have anything to do with

1:18:56GitHub. So like for example, if I want

1:18:58to create a PRD, I can say, you know,

1:19:00use Arkon to walk me through creating a

1:19:02PRD. And so this is going to uh you

1:19:05know, hopefully load the Arkon skill

1:19:07here and then walk me through the

1:19:08interactive PRD workflow. So I'm just

1:19:10showing off another workflow really

1:19:12quickly. And this one's actually pretty

1:19:13cool cuz it has human in the loop. So it

1:19:16asks you some questions. It starts

1:19:17building the PRD and then it stops to

1:19:19ask you more questions. So maybe we'll

1:19:22see this in action really quick here.

1:19:24And then I also I do want to get into

1:19:26the other

1:19:28um

1:19:29the other thing I wanted to show you

1:19:30guys like building a workflow based on

1:19:32GSD.

1:19:34Okay.

1:19:36Um pick your starting points. Let's see.

1:19:40Okay, so it actually tells me like

1:19:41here's a few different workflows. Um I

1:19:43want to use the interactive PRD

1:19:45workflow.

1:19:47So, let's have it kick that off for us

1:19:49here.

1:19:53Okay.

1:19:55Interactive workflows.

1:19:58Uh got it. The interactive PRD workflow

1:20:00is a guided conversation where AI asks

1:20:02questions in rounds to build out a PRD.

1:20:05Cool. So,

1:20:07um

1:20:07I don't know.

1:20:08I want to build support for Pi agents in

1:20:11Arkon. Well, that I mean I think it's

1:20:12kind of an interesting example that you

1:20:14know, let me make sure actually my

1:20:16speech-to-text tool spells that right.

1:20:18And I'll just say uh with the SDK. Not

1:20:22that I'll build this right now, but I'll

1:20:23I'll show you guys quickly what it looks

1:20:25like to go through a different workflow

1:20:26cuz like I get so hyper-focused on using

1:20:29the fixed GitHub issue workflow. I want

1:20:30to show you guys something a little bit

1:20:32different here.

1:20:33Okay, so there we go. So, now it kicks

1:20:35off the Arkon investigate or sorry,

1:20:37interactive PRD. So, the workflow is

1:20:39running. It's going to explore the code

1:20:41base before it starts asking us some

1:20:43questions here.

1:20:44And we can obviously go and like view it

1:20:46in the logs along with the other ones.

1:20:48So, this workflow is actually in a

1:20:50paused state now. So, you can see in the

1:20:51web UI we have support for this where

1:20:52it'll show us like here is where we have

1:20:55a human-in-the-loop step. So, while the

1:20:57all the other workflows are running,

1:20:58this one we're more interacting with.

1:21:00So,

1:21:01uh let me So, it says it's paused. Let

1:21:02me grab the output. So, it's going to

1:21:03read the logs and it's going to uh ask

1:21:06us some questions here, right? So, we

1:21:08the workflow is in a paused state for us

1:21:10to give feedback. And this is very

1:21:12similar to what I was showing in the uh

1:21:14example in the diagram earlier where we

1:21:16could have like human-in-the-loop for a

1:21:17planning step.

1:21:19But, we get to actually review the plan

1:21:20and have the coding agent iterate on

1:21:22this before we go to the next step. So,

1:21:25asking some foundation questions.

1:21:27Uh let's see. So, who has this problem?

1:21:31Um

1:21:32everyone using Arkon. I think it's kind

1:21:33of a weird question actually. What

1:21:35problems are they facing?

1:21:37Uh the problem they're facing right now

1:21:39is there a lot of people that don't want

1:21:40to use Claude or Codex specifically.

1:21:42They want to use different models. Pi is

1:21:44more of a general agent that makes it

1:21:45really easy

1:21:47to uh to use other models.

1:21:50Um

1:21:51let's see. Why can't they solve it

1:21:53today? Uh well, we don't have support

1:21:55for Pi yet. And I want to build this now

1:21:57because we're currently working on

1:21:59making big improvements to Arkon and Pi

1:22:01is one of our priorities. We will know

1:22:03when it's solved. Success looks like we

1:22:06can use Pi any model in Pi with all of

1:22:09our Arkon workflows.

1:22:12So, these questions um if you aren't

1:22:13familiar, these are like really standard

1:22:15questions for product managers to ask

1:22:17when they're first creating a PRD. So,

1:22:19we have a like a lot of like product

1:22:20manager best practices built into this

1:22:23workflow.

1:22:25Um now, the some of the questions were a

1:22:27little awkward for me right now just for

1:22:29the demonstration here, but they they

1:22:31are like legitimately good questions to

1:22:33ask when you're first creating a product

1:22:34requirements document. So, now you can

1:22:36see that um using the Arkon skill, Arkon

1:22:39knows like, "Okay, let's resume the

1:22:41workflow within approve state." So,

1:22:43we're approving and then it's giving our

1:22:45feedback in.

1:22:47So, it passed our answers through. So,

1:22:49now the workflow is it went out of the

1:22:51paused state into the running state. We

1:22:53can of course view the logs here to see

1:22:55what's going on.

1:22:57So, this is the first node that it ran.

1:23:01And then it asked us some questions.

1:23:02Now, we move on to the next one. So,

1:23:03it's kind of just like a process of like

1:23:05asking us questions and going in a loop

1:23:07here.

1:23:10All right.

1:23:11Um oh, is it paused again? Hold on. No,

1:23:13it's still running. Okay. So, we'll wait

1:23:15for it again.

1:23:17Although, I might actually not continue

1:23:19with this demonstration cuz I think you

1:23:20guys get the idea. It really just does

1:23:23this in a loop where it has like a set

1:23:24of questions that are it goes through

1:23:26for each one of the the prompts that we

1:23:28have right here.

1:23:30So, I think you guys get the point. But,

1:23:31there's an example of a workflow where

1:23:34um it doesn't really have anything to do

1:23:36with GitHub cuz not all of them have to.

1:23:38That's what I wanted to show. The final

1:23:40artifact obviously of this workflow is

1:23:42going to be a markdown document which is

1:23:44our PRD. So, then we would, you know,

1:23:46break that up into tasks and and go

1:23:48through a PID loop as a separate Arkon

1:23:50workflow to uh knock out all the phases

1:23:53that we have in that PRD.

1:23:55And then if you really wanted to get

1:23:56fancy, you could even if you wanted to

1:23:59make it so that the whole like PID loop

1:24:01process and PRD creation like everything

1:24:04is a single Arkon workflow. Like I said,

1:24:06your entire software development life

1:24:08cycle, you can package that all up as an

1:24:11Arkon workflow.

1:24:15All right.

1:24:17Cool.

1:24:18So, yeah, I think with that I want to

1:24:20show you guys how to create a workflow

1:24:23from scratch.

1:24:25Cuz it's actually beautifully easy.

1:24:28So, I'm going to open up another Arkon

1:24:29session here. And you know, before I do

1:24:31that, I do want to spend some more time

1:24:33with Q&A with you guys. So, let me go

1:24:36back to the full frame here.

1:24:39And I will open up some more questions.

1:24:44All right.

1:24:45So, give me a second to kind of read

1:24:47through what we got here.

1:24:48And then I think uh yeah, so Rasmus is

1:24:51is here in the live stream now answering

1:24:52questions. Appreciate it a lot, Rasmus.

1:24:55And then thank you, Thomas, as well.

1:24:59Um

1:25:00let's see.

1:25:07Uh what is the simplest path to get this

1:25:08to work with linear tasks instead of

1:25:11GitHub issues?

1:25:13Yeah, so my recommendation would be to

1:25:15use either the linear MCP or create a

1:25:18skill to use the linear API.

1:25:21And then you would build a custom

1:25:22workflow.

1:25:23So, instead of like fixed GitHub issue,

1:25:25it would be like fixed linear issue or

1:25:28like handle linear task. And you could

1:25:31even have Arkon reference the fixed

1:25:34GitHub issue workflow

1:25:37and use that as a starting point where

1:25:38you would just mold things to be all

1:25:40linear instead of all GitHub.

1:25:43So, it's actually very easy to

1:25:46integrate any platform that you want

1:25:47into Arkon because for every single

1:25:50node, you can inject skills or MCP

1:25:52servers.

1:25:54Um so, for the case of GitHub like

1:25:56coding agents are so good at using the

1:25:58GitHub CLI that I don't need a GitHub

1:26:00CLI skill. You might need that for

1:26:02linear though or like I said, the MCP

1:26:04server.

1:26:06But seriously, all you have to do and

1:26:08I'll show you this in a second is you

1:26:09like okay, watch this.

1:26:11Um

1:26:12I don't know I don't think I'll actually

1:26:13run this right now, but I Oh, here. Let

1:26:15me go back to my scene so you guys can

1:26:16see my screen. I forgot I wasn't sharing

1:26:18my screen. So, I can say like

1:26:21uh load the Arkon skill. I want to

1:26:23create a version of the GitHub fixed

1:26:25issue workflow, but specifically for

1:26:27linear instead of GitHub.

1:26:30Like that is all you have to do. Now,

1:26:32obviously, you're going to probably have

1:26:33to iterate quite a bit on the prompting

1:26:35and like really making things specific

1:26:37to you. Uh so, this is this is certainly

1:26:39an oversimplification, but like this

1:26:41literally could be your starting point.

1:26:43Like this is what you you send in.

1:26:45Um and so, yeah, after a little bit of

1:26:47Q&A here, I'll show you what it looks

1:26:49like to create a workflow in like Man,

1:26:51the Arkon skill is so beautiful because

1:26:53it just it knows everything. It'll It'll

1:26:55walk you through everything. And And

1:26:56another thing I love doing with coding

1:26:58agents is

1:27:00I like asking it to ask me questions.

1:27:02And so, maybe I would send in Let me go

1:27:04back to my aqua voice here.

1:27:08So, I can paste this back in. Um and I

1:27:09can say like, you know, ask me questions

1:27:11to make sure you understand uh my linear

1:27:13setup and exactly how I want the

1:27:16workflow to function. Right? So, that

1:27:18way you're you're kind of reducing the

1:27:19assumptions that it makes up front.

1:27:23Um and so, what it produces in the end

1:27:25is going to be better aligned with what

1:27:26you actually want before you go into

1:27:28iterating. You'll probably still have to

1:27:29iterate, but that's the idea here is

1:27:31like go through a bit of a planning

1:27:32process with Arkon

1:27:34before it um

1:27:36before it creates the workflow for us.

1:27:42All right.

1:27:46Cool. What else do we got here?

1:27:50>> [snorts]

1:27:50>> Um with the session management within

1:27:52the workflows, how is context

1:27:54persistence handled? Is the data and

1:27:57context passed through the nodes?

1:28:00Okay, that is a really good question.

1:28:02And uh we have a few different ways to

1:28:04handle context persistence here.

1:28:08So,

1:28:09uh I'm trying to think if I have a good

1:28:11workflow to demonstrate this. Instead of

1:28:14poking around, I might just more answer

1:28:15your question at a high level.

1:28:17Um but yeah, one [clears throat] of the

1:28:19core things that we have in Arkon is a

1:28:21parameter for each node that specifies

1:28:24if we want to continue the session from

1:28:27the prior node or start fresh.

1:28:30And so, that that flexibility is pretty

1:28:32powerful because maybe you want a

1:28:34different node where you inject

1:28:36different skills

1:28:37or you just want to like have a new node

1:28:39for some reason cuz you want to like

1:28:41switch models or something, but you want

1:28:43to still continue the same conversation

1:28:45from the previous node. You can do that.

1:28:47Or you can say like I actually want to

1:28:49start completely from scratch in this

1:28:51node.

1:28:53And then another thing that we have in

1:28:54Arkon

1:28:55is we have like outputs. Like we can

1:28:57output artifacts.

1:28:59And so like for in this case, we have

1:29:01the planning step

1:29:04and it's going to output a plan to our

1:29:06artifact directory for our current

1:29:08workflow execution. So, this is like one

1:29:10of the primitives we have in workflows.

1:29:12It's like the artifact dir.

1:29:14And then we when we go in a brand new

1:29:17session in the implement stage,

1:29:19we are going to prompt it to read the

1:29:22plan from the artifact directory. So, if

1:29:24I look at the Archon fix issue command,

1:29:27let me open this up here cuz we you can

1:29:29prompt inline. I've shown that in some

1:29:32of the workflows. You can you can have

1:29:33the prompt right in the YAML for the

1:29:35workflow. Or if you want to and you want

1:29:37it to like kind of be more organized,

1:29:39you can have it reference a command in

1:29:41the commands folder. So, we'll go to

1:29:43Archon

1:29:45fix issue. There's a lot of commands

1:29:47here, so I have to find the right one.

1:29:50There we go.

1:29:51So, Archon fix issue, you'll see here

1:29:53that the argument is the artifacts path.

1:29:57So, Archon is smart enough to know like

1:29:59once we are done with the planning and

1:30:01we have our plan here, we're going to

1:30:04prompt in a brand new session for our

1:30:06coding agent to read that plan.

1:30:09Right? Like go to the artifact directory

1:30:11and read that plan and then go through

1:30:13the implementation here. So, it's a

1:30:15brand new session, but then we're still

1:30:17passing some context or like I said, you

1:30:20can just

1:30:21have the context be like continue

1:30:22instead of fresh.

1:30:25So, it's up to you like flexibility

1:30:26depending on how you typically work.

1:30:31All right.

1:30:38Let's see what else we got here.

1:30:43Let's see. Have you really given

1:30:44anything other than Opus a real try at

1:30:46agentic coding tasks? I haven't.

1:30:49Uh so, I I do mostly use Opus for my

1:30:51implementation. However, I have tested a

1:30:54lot of using Sonnet for my Archon

1:30:56workflows.

1:30:57So, I talked about this a little bit at

1:30:58the start of the live stream, but I get

1:31:00better results using Sonnet to fix

1:31:03GitHub issues with this workflow than

1:31:05just using Opus by itself in Claude

1:31:07code.

1:31:09Because of all of the context

1:31:11engineering that goes into you know,

1:31:13like this I guess you could call this

1:31:14like a GitHub issue harness, like a

1:31:16fixer harness.

1:31:18Uh so, I have run this with Sonnet. So,

1:31:21Sonnet is actually the default model for

1:31:23all the nodes here. Uh now for the

1:31:25implementation itself, I think we have

1:31:27it set to Yeah, so we have it set to

1:31:29using Opus. So, like the default

1:31:31workflow as it stands uses Opus

1:31:33specifically for implementation, but

1:31:35obviously, if you just delete this,

1:31:37it'll use Sonnet as the default or um

1:31:40you know, you can change the model

1:31:42manually or whatever. But let's say like

1:31:44at one point I had it set to Sonnet and

1:31:46I was using it Sonnet for everything.

1:31:48Sonnet or Haiku for everything here and

1:31:50I was still getting really good results.

1:31:51Obviously, obviously, the best model is

1:31:54always going to give you the best

1:31:55results.

1:31:56Um but for the sake of like making sure

1:31:58you don't hit your rate limits and

1:31:59stuff, you can use Sonnet for these

1:32:01workflows.

1:32:02And you can even like ask it to adjust

1:32:05the model at the default level or the

1:32:07specific node level before you run it if

1:32:09you want to.

1:32:13All right.

1:32:15People like seeing you failing. Pretty

1:32:17sure it makes you a normal human. I

1:32:19assume you referenced that when I was

1:32:21having all of the

1:32:22issues here in the VPS. So, I appreciate

1:32:24that. We'll talk about why I have those

1:32:27issues in a second cuz I have a Claude

1:32:29code I'm not actually authenticated with

1:32:31Claude.

1:32:32I am authenticated with the Minimax API.

1:32:35That's what was causing the problems on

1:32:36the VPS cuz I'm getting things set up

1:32:38for the Dark Factory.

1:32:41All right. How many tokens does start

1:32:43the app take?

1:32:45So, when I asked it to just like spin up

1:32:47the front end and back end of Archon,

1:32:49I mean, I don't know exactly, but it's

1:32:51probably just going to be a couple

1:32:52hundred unless it's like running into an

1:32:55issue for some reason. It's not going to

1:32:56take that many tokens.

1:32:59And by the way, you can see here that

1:33:01it's already kicking off the whole

1:33:02process here helping me create the

1:33:05linear version of the workflow. So,

1:33:06asking me some questions here. I'll

1:33:08answer those and then it'll create the

1:33:10workflow and test it for me and

1:33:11everything. Like I won't go through this

1:33:13full process here, but for the question

1:33:15earlier on on migrating to linear, like

1:33:17it's just so easy to have it walk

1:33:19through cuz pretty much like here's the

1:33:21thing. All of the default workflows that

1:33:23we have in Archon,

1:33:25there's two there's two uses for them.

1:33:27One is you can just use them directly

1:33:29out of the box if if there is one that

1:33:31like matches how you already work, but

1:33:33the other maybe even like more important

1:33:35part of these default workflows is it's

1:33:37a reference point for your coding agent

1:33:40to build something that's actually

1:33:41custom to you.

1:33:43And so, even if you're building

1:33:44something that's like super different

1:33:46than everything we have here, like some

1:33:47kind of like refactoring workflow for

1:33:49example. Actually, we have that as well.

1:33:51So, maybe that's not the best example,

1:33:53but even if you have some like super

1:33:55different workflow that's like not even

1:33:57close to anything we have here,

1:33:59you still want your coding agent like

1:34:01loaded in the Archon repo to look at

1:34:04these as a reference to understand like

1:34:05the different parameters we have for

1:34:06nodes and how we handle loops or

1:34:08whatever else like deterministic nodes,

1:34:10whatever you might want in your

1:34:11workflow.

1:34:13So, very easy to pretty much build any

1:34:15I'll show that in a second, but I just

1:34:16want to answer a couple more questions

1:34:17here cuz you guys have a lot of really

1:34:19good ones.

1:34:21All right.

1:34:24Let's see.

1:34:26All right. [snorts]

1:34:30Yep, Pi next week. Yeah, not not a

1:34:31promise that like we'll have it ready

1:34:33ready next week, but I'm definitely

1:34:35going to be working on me or Rasmus will

1:34:37be working on the pull request for

1:34:38adding in Pi.

1:34:42Um could I make a node to use Codex

1:34:44Spark in a workflow?

1:34:47Um I mean, there's nothing stopping you.

1:34:49You can make a node do literally

1:34:51anything because it can be a prompt to a

1:34:53coding agent or it can be a script. Even

1:34:55if you wanted to use a coding agent

1:34:57that's not supported by Archon directly,

1:35:01you can just make a script that invokes

1:35:02that. Like the Yeah, the world's your

1:35:04oyster here.

1:35:08All right.

1:35:12Um let's see. What about

1:35:14create a node to mention other workflow

1:35:16not only a command? I I I think what