AI Agents Part 2: Multi-Agent Orchestration

0:04All right, welcome everybody to part two

0:07of the age of agents workshop. This week

0:09will

0:10today we'll be talking about multiple

0:13agents and how we can orchestrate them

0:14all together. My name's Mike Schwartz.

0:16I've been running these sorts of

0:18workshops for about four or five years

0:20now. I've been working in technical

0:21automation for 26 years since I

0:23graduated from UBC in 2000. Just been

0:26building technical automation projects

0:28and deep into AI for the last

0:30five years. So for those of you who did

0:34not attend the first workshop, you were

0:35supposed to and hopefully did your

0:36homework and watch that video.

0:39The first workshop will do just like a

0:40little a quick recap on what that one

0:43was about.

0:44But today is all about moving from

0:47single agents into multiple agents and

0:49really proving the point in like what

0:51are the situations when more is better.

0:54I think most of those most of the time

0:57that is the case. So the previous

0:58workshop was all about like the agents

1:00have arrived. There's an absolute

1:02explosion of agents right now. Our

1:05agency is just crushed with demand.

1:07Anybody that is

1:09selling agents as a service right now,

1:10we've just seen going from crickets in

1:13December, January to just now it's just

1:15an explosion. Everyone's panicking. The

1:17entrepreneurial world right now,

1:19everyone's panicking about agents and

1:20that can be really reflected in revenue

1:22growth over at Claude.

1:25Look at this chart. It's straight out of

1:27the slide.

1:28Pretty crazy growth. Now the crazy part

1:31is that from March until today, they

1:33increased like they doubled again.

1:35They're over 40 billion dollars in

1:38ARR. So just meteoric growth and this is

1:40just one of the language models that is

1:43powering the agentic revolution.

1:45Probably one of the largest, but the

1:47growth is absolutely insane. And this is

1:50why Big Tech is spending so much money

1:53investing in data centers and chips is

1:57because agents they use a heck of a lot

1:59more compute than normal AI. And the

2:04vision that they have and we'll see if

2:05the stock market will keep its lofty

2:07heights is that all of this compute

2:09capacity is going to be gobbled up by

2:11entrepreneurs like us. People who

2:13realize that hey, I can build a

2:17instead of having a 10 person marketing

2:18team, I can have one expert that

2:20controls an army of agents and now

2:22they're running my marketing department.

2:23We're seeing companies shrinking in head

2:26count and amplifying and increasing the

2:28number of agents that they have

2:31working for them. So that's what all Big

2:33Tech is betting on. All the venture

2:35capital money is betting on this and

2:37right now so far, if you look at

2:39Google's share price this morning, all

2:42time high. Nvidia yesterday, all time

2:44high. The trend is continuing.

2:47And we're on the edge of this

2:48intelligence explosion. So there's never

2:50been a better time than now to move fast

2:53into this world. It is only going to

2:55accelerate from here.

2:57So that was a little bit from the the

2:59last workshop.

3:02Now

3:04I hear a lot from people that you know,

3:06you've probably heard me say before the

3:07term like drinking from the AI fire

3:09hose. Again, these talks on their own

3:12will not be enough for you to fully nail

3:15these concepts. It's meant to sort of

3:17let you know what the important things

3:18are so that you can take them offline

3:20and study them. We could pick any one of

3:22these slides and probably spend an hour

3:24talking about it. So keep that in mind.

3:26We'll send you the recording. Janette

3:28will be sending out some homework and

3:29links.

3:30And there's lots of other workshops. So

3:32next month we'll have a workshop, two of

3:34them. One on how to secure your agents,

3:36another one on how to make your agents

3:38cost effective. We call it token

3:39optimization. So there's lots of

3:41different topics. So don't feel

3:44too much pressure to catch everything

3:47right away. This is a marathon journey.

3:49So I thought we'd start this out by just

3:51breaking down some of the elements of

3:53like what makes up an agent. And there

3:57are a lot of different elements and

4:00I've broken them down into four key

4:02categories. So the first thing that an

4:03agent needs to have is access to a large

4:05language model to get its tokens. Tokens

4:07are really the the unit of exchange, the

4:09way that we

4:11measure cost. So in the past when you

4:13would connect to a server, we would pay

4:15for bandwidth. Remember when bandwidth

4:17was expensive? Think back 10, 15 years

4:19ago.

4:20Now bandwidth is pretty much free. We

4:22don't think about it too much anymore.

4:24But as we interact with agents, we have

4:25to pay for that intelligence. It comes

4:27back and forth to us in the form of

4:29tokens, which are really just chunks of

4:31words and think of it as the unit of

4:32exchange for purchasing

4:35intelligence from an LLM.

4:38And I've got a live demo here from our

4:42multi-agent orchestration platform, but

4:43this these concepts can work for any

4:46platform whether you're using

4:49Open Claw,

4:50the new Hermes open source agent, which

4:53is blowing up. That's a free one.

4:55Paperclip.

4:57Some direct competitors to us would be

4:59like Perplexity computer,

5:01Claude agent teams and many more. This

5:04is probably I think going to be the

5:06fastest emerging business model of 2026,

5:09multi-agent orchestration platforms. So

5:11when I talk about what we're doing, I

5:13want you to know there is there's a lot

5:14of others out there. This is not

5:17you have a lot of options and we'll go

5:18through sort of choosing the right uh uh

5:21path for you later on.

5:23So let me show you how that what that

5:25looks like

5:26within our platform.

5:29So first of all, credentials. On the

5:31back end, this is the the control room

5:33for our agent system.

5:36So

5:37what we do, we just connect directly

5:39with Claude for our tokens, for the the

5:43main intelligence tokens. Now

5:46through other credentials, we also

5:48connect to things like Open AI for

5:51certain types of intelligence or I might

5:53connect to Gemini to get images from

5:55Nano Banana.

5:57I'm connecting to 11 Labs to get

6:01voice for my voice agents when I call in

6:04over the phone. So there's different

6:05agents different LLMs that we connect to

6:09for different purposes. But for us, the

6:11core system intelligence is all coming

6:14from Claude. Right now, Claude Opus 4.7

6:18seems to be winning that game. Now GPT

6:215.5 Codex just came out a few days ago.

6:25I haven't tested it yet, but I've heard

6:26some really good things about it. So

6:28there is a chance that ChatGPT just went

6:30from second place back to first place.

6:32We'll find out in the next few days as

6:33we get deeper into testing GPT 5.5

6:36Codex.

6:37But just back to the the elements of an

6:40agent, your agents need to connect to

6:42intelligence. That intelligence is

6:43coming from different language models.

6:46You'll typically have specialist models

6:48like we use Gemini for images and video,

6:5111 Labs for voice. We might use

6:53Perplexity as

6:56for deep research.

6:58But you're going to choose probably 90%

7:00of your work will be done by one model.

7:02For us, that's Claude Opus 4.7.

7:05And we connect and you heard last time I

7:08to be careful what I'm saying here

7:09putting this out on the public web, but

7:11there's a short-term hack

7:13that a lot of people are doing when you

7:15can connect to your Claude accounts via

7:17the subscription layer

7:19versus not paying through the API.

7:22Take that offline, do some research, ask

7:24some agents, but there are some very

7:26cost effective ways of carefully

7:28connecting your agents to these models.

7:30So

7:31um

7:33tokens. Your agents need to connect to

7:36LLMs.

7:37The next thing agents need to do is they

7:39need to have access to all of your data,

7:41all of your systems.

7:44And we call this on our side the

7:47credential connection. So what that

7:49looks like for me beyond the LLMs are

7:52what are all of the different data

7:54sources? Say looking at my company here,

7:56Myzone AI. What are all the different

7:57data sources that I want my agents to

8:00have access to? So this is beyond the

8:02LLMs. This is things like well, I have

8:05all my tasks in Asana or I have all my

8:07emails in Gmail.

8:10There's a long list of data connectors.

8:13I can show you this list here.

8:17So we have a an onboarding

8:20component that helps us sort of confirm

8:22what all those data connections are for

8:24for each of our clients. And I'll sort

8:25of walk you through the list of the most

8:27common. But if we say start with the

8:30communication layer,

8:32companies need to connect either Slack

8:34or Microsoft Teams or under productivity

8:37suites are using Microsoft 365 or Google

8:40Workspace.

8:41What are you using for your CRM?

8:42HubSpot, Salesforce, Pipedrive, custom

8:45or other? Project management software?

8:48Video conferencing software?

8:51Are you using any AI note takers?

8:53Plaud, Pendents, Fireflies. We saw a few

8:57hundred note takers try to pop into this

8:59meeting. Even your phone systems. If

9:01you're on a VOIP phone system,

9:04most of them like RingCentral, the one

9:06here in Canada, Telus is powered by

9:07RingCentral. It it's VOIP. It has an

9:10API. You can connect into that. And

9:13there's a long list.

9:16Help desk support, scheduling software,

9:18accounting software, e-commerce, your

9:20websites.

9:21And many many more. So

9:24most people start with maybe 10, 15

9:27credentials, but over time your list of

9:29credentials will start to look like

9:31this.

9:33Now this is a little bit of OCD

9:35>> [laughter]

9:36>> working on agents.

9:37But I'm I'm connecting into absolutely

9:39everything. Different hosting accounts.

9:41I can even connect into GoDaddy now to

9:44manage my DNS or like subdomains. I

9:48recently just reconnected into Amazon

9:52Route 53 so we can manage all of our our

9:54DNS settings across all of our client

9:56accounts,

9:57BrowserStack for mobile usability

9:59testing, and more. Now, this is a very

10:01complex list. Most companies are not

10:03that deep on credentials.

10:06But okay, so now your agent has access

10:08to LLMs for intelligence, it has access

10:10to all your data systems. That also is a

10:14combination of data plus tools. So,

10:17there can be additional tools that you

10:19add credentials to. An example of a tool

10:21that you can add to your agent through

10:23these credential layers would be

10:27BrowserStack. BrowserStack, I use that

10:29tool to enhance my agents with mobile

10:32usability testing. So, I can emulate an

10:35iPhone

10:361 through 17 and see what it looks like

10:39when I'm doing usability testing, for

10:40example. So, think of that credential

10:43layer. If you're in say Claude co-work,

10:45a lot of this, I think they refer to it

10:47there as connectors. There's different

10:49names for it with different agent

10:50models.

10:52Now, once you have all of your

10:54credential set,

10:56you need to set your data platform, your

10:59data layer. We internally refer to this

11:03as the brain.

11:04Now,

11:07the brain

11:08looks

11:10well,

11:13did my onboarding. Here we go.

11:15So, the company brain is the next layer

11:18up when we're setting somebody up. And

11:22what we do there is if you think of all

11:24of your

11:26data systems, think of it as like

11:30uh the brain, think of it as like a

11:31local cache of all these different data

11:33systems. Because if you don't have local

11:36accessible data for your agents, what

11:38ends up happening, I'll give you a lot a

11:40a real example that happened with me if

11:42a month ago when I realized this.

11:44Imagine I had my agents that were

11:46connected to

11:48my Google Drive for all of my SOPs.

11:51Well, if I'm asking my agent

11:53to give me a summary of something that's

11:55in my SOP, the agent would have to go to

11:58Google Drive,

11:59download all that information, put it

12:01into context window, use a bunch of

12:03tokens, probably spend 3 4 minutes, and

12:06then give me an answer.

12:08Now, the way the company brain works

12:10within an agent model is it will

12:11localize all that data, will sync it up

12:14with a series of jobs.

12:16So, say for example, you want your brain

12:18to contain all client inform-

12:20client information across CRM, task

12:23management, Slack or Teams,

12:26vo- phone calls, every- all your company

12:28communication should come into your

12:31brain.

12:32And so, we set up these little

12:33synchronization jobs that will go and on

12:36different time intervals.

12:38Um I'll I'll visually illustrate this

12:39here. We call these jobs.

12:42Uh these are in old school programming

12:44world that these are cron jobs.

12:47Um not like employment jobs, those are

12:48rapidly disappearing.

12:50And these jobs, what they'll do is will

12:52run different syncs. We call them like

12:54the brain sync to Asana, to Gmail, to

12:56Slack, to different

12:58data sources.

13:00Right? So, we have the the agents will

13:02go out, sync, collect it, organize it

13:04locally in your brain.

13:06And

13:08what this allows is

13:10way faster, way cheaper retrieval of

13:12information

13:14uh for your agents.

13:16So, that's the third and important

13:18layer. Now, once you've got your

13:19connected to all your LLMs, connected to

13:21all of your credentials, you've got all

13:23of your data synced up and localized in

13:25your own private server, either a Mac

13:28mini

13:29or a virtual private server up in the

13:30cloud. This is all for security reasons.

13:32We we need separate hardware for agents.

13:34We don't trust them on our own

13:35computers.

13:37And we've got that data localized and

13:39organized, then the next layer we get

13:41into with agents are building skills.

13:44Now, skills, I think this is Claude's

13:47fancy way of

13:49I mean, in the past they were just known

13:50as prompts. ChatGPT referred to them as

13:53custom GPTs. Um but skills are really

13:56just detailed prompts that give you

13:59step-by-step instructions on what an

14:01agent is supposed to do for uh specific

14:04things. I think of the skills as little

14:06LEGO pieces, building blocks that we can

14:08snap together to create some pretty

14:11magical things.

14:13So, in these individual elements,

14:14nothing here is like super exciting at.

14:16The magic is is when we start to snap

14:18them together.

14:19So, I'll give you some illustrations of

14:20skills or those LEGO pieces that we use

14:23to build our agents.

14:25So, if I go over here to my skills,

14:29you'll see a long list of skills. A

14:31skill to add to brain or to optimize for

14:34answer engines at like LLMs for SEO or

14:37AEO,

14:38a skill to access a browser as an agent,

14:42or a skill to create an SEO report,

14:46a skill to properly connect and talk to

14:49my Slack channels,

14:50a skill to manage my DNS with Route 53

14:54from Amazon,

14:55and so on and so forth. A long list.

15:00Now,

15:01what is an agent? An agent is really

15:03when we wrap all of these things

15:05together, we put a nice little hat on it

15:06and say, "You are a security agent."

15:10So, for example, we created a whole

15:12bunch of different skills related to

15:14security. As we know with agents and

15:16what next month we have a whole talk 90

15:18minutes dedicated to security. Very,

15:21very important for agents.

15:23So, we built and we researched a whole

15:25bunch of different skills related to

15:26security that agents would need.

15:29Um and

15:32let's look at one of them.

15:35Security

15:38metrics.

15:39This skill that was researched is

15:42there's about I think 15 different

15:44skills that roll up into our security

15:46agent. And a skill is really just a

15:48prompt. If you read down on what that

15:52individual skill does, it'll just look

15:54like a prompt that you've probably seen

15:55elsewhere.

15:59Example outputs.

16:02We won't need to get too much into the

16:03detail here. But once you've got all of

16:05your different skills related to an

16:07agent,

16:09then we roll that up into an agent. And

16:13these are all the different agents that

16:14I have. These are a collection of

16:15skills, connectors, credentials, tools,

16:19data access, and we we put a a fancy hat

16:22on them, give them a name, and then we

16:24can call them very quickly. So, if I go

16:25back to that security agent, my security

16:28agent

16:30is

16:31really it's a set of instructions, and

16:33it says you've got access to all of

16:35these individual skills.

16:38I also connect them to some dynamic

16:40workflows, which we'll show you in a

16:41second.

16:43And it has typically

16:45an an initial in- set of instructions.

16:48We call this the agent or agents.md

16:51file. Every agent will have a set of

16:54instructions that says, "You are a

16:56security agent. These are your roles.

16:57These are the things that you're going

16:58to do. These are all the skills that you

16:59have access to."

17:01So, we're building up. This complexity

17:04builds into simplicity. So, that's the

17:06good news for those of you that just

17:07want to use agents and don't care about

17:09all these building blocks.

17:11So, here you can see from the agent.md

17:13file, this is the overview so that the

17:15agent knows who they are, what they're

17:18supposed to do, and how it all works.

17:19So, this agent has three hats, red hat,

17:21blue hat, auditor mode.

17:26And it has access to all these skills.

17:28Again, this is a high-level quick

17:30review.

17:31Now, if I just want to use an agent,

17:34this is where things get

17:36easy.

17:38So, say I want to call in a security

17:41agent.

17:45Now,

17:46have any volunteers um that want to post

17:48just the first person in the chat window

17:50give me your domain name, we'll do a

17:51security review of your website, and

17:54then Janette, you can just read out the

17:55domain name for me.

17:58We got one really fast.

18:00>> [laughter]

18:00>> Lori Fencing. No way, sorry. Uh Fence

18:03More. Uh Christine Cruel. So, f e n c e

18:08n m o r e.com.

18:15fencenmore.com.

18:17That was fast.

18:19Um

18:21can you please run a detailed security

18:23review on this website? I want the

18:26output to be a beautiful interactive

18:28HTML website. Really impress the client

18:31here.

18:33Make it on a publicly accessible URL.

18:38And

18:39honestly, some of these agents, all I

18:40have to do is just put in a a URL and

18:42hit enter, and it will already know what

18:44to do because Hey, you're missing an n.

18:49It's f e

18:51n c e n

18:53m o r e.

18:56fencenmore. So, this has nothing to do

18:57with fences.

18:59>> [laughter]

19:00>> I thought this was somebody selling lots

19:01of fences. Okay, fencenmore.com.

19:05All right.

19:06So, we'll kick that one off. Um you saw

19:08that last talk, we did one for SEO

19:10agent, CRO agent. But so, once you have

19:13all of those connectors and credentials,

19:15you've got your tokens, you've got your

19:17skills, they're all rolled up into an

19:18agent. The agent is told, "Act in this

19:20role. This is how you behave." You just

19:22come along and tell the agent what you

19:24need to do.

19:25And

19:27okay, yeah, I'm seeing the note there. I

19:28was spelling that right. And so, there's

19:30a lot of complexity behind the scenes,

19:32but this is where it gets simple is no

19:34matter what you need. If you've got an

19:36agent that's built for that, like

19:39let's get um maybe another

19:42s- uh domain. I saw Jason posted there

19:45printprint.ca. Let's

19:47let's add that one. I'll grab now a

19:49design agent, which has rolled up a

19:52whole bunch of different design-related

19:53skills and say,

19:56"Can you

19:58please visit, analyze, audit

20:00printprint.ca

20:02website, and I want you to

20:05mock up a new improved site design for

20:08them that will improve their conversion

20:10rates."

20:13Right? So, it's

20:15what do you need the agent for? We'll

20:16have specialized agents for for

20:18everything. Here's a list of some of my

20:20agents. I've got

20:21agents that are handling management of

20:23all of my data and data systems, the

20:25brain manager. So, I can say things like

20:28"Connect to my QuickBooks and

20:31synchronize that on a daily basis to

20:33ensure all of my key financial data is

20:36accessible to my leadership team and

20:38organize that in my private

20:41brain."

20:42And it will it will do that. Or

20:45communications agent that can handle

20:47reading, drafting emails, sending out

20:50various communications, design agents,

20:52developer agents, marketing agents,

20:55onboard This is to help me onboard new

20:57clients more efficiently,

20:59and so forth. So,

21:02that's where things start to get magical

21:04for people is

21:06when you can build all this behind the

21:08scenes. Like I I built one agent it was

21:10called a landing page agent. I gave

21:13access to my team through a Slack

21:15channel, and now they just go into Slack

21:17and say, "I need a new landing page for

21:18Mike's next workshop."

21:20And in two, three minutes, they have

21:22their their landing page

21:24done.

21:26So, those are some of the elements.

21:27Tokens, credentials, data layer, skills,

21:28and all of this gets rolled up

21:32to

21:33the agent.

21:34So, we went through uh tokens,

21:36credentials, data layer, skills. Okay,

21:38so the agent is really just bundling

21:40that all up together. You just call it

21:42by name, and instantly you have access

21:44to skills, data, credentials, and tokens

21:46in one shot. Um that is the agent. Now,

21:50there's some real difference in how we

21:53use agents.

21:55Either you can have like a single agent

21:57for a task, like you just saw me do,

21:59like, "Hey agent, go do this."

22:02Um or you can also have agents work in

22:03these more complex workflows. I call it

22:05like a single agent workflow. If you're

22:07using Claude CoWork as your first agent

22:09experience, which is a great We

22:10recommend most of our clients start

22:13learning about agents by using Claude

22:14CoWork. Um and then when you're ready to

22:17move into multi-agent orchestration,

22:19then you'd get into like Open Claw,

22:21Hermes, Paperclip, Myzone, AI 1, um

22:25Perplexity computer, or others. So, but

22:27here we're going to look at what a

22:28single agent uh workflow looks like. And

22:31this is what you would really experience

22:33on Claude. This is where you have

22:36really is one agent that's just

22:37switching hats.

22:39Right? And there's pros and cons to

22:41that. I mean, the pro is it's quick,

22:42easy to do. Cons, it can be a bit more

22:45expensive to run from a token

22:46perspective.

22:47But this is one agent that just switches

22:49between roles and will work autonomously

22:52on a long set of of tasks. So,

22:54here's a I'll give you example of a

22:57a process or we call these recipes on

23:00our platform, um using the blog bot. And

23:03this is kind of what the the publishing

23:05pipeline would look like. We've got one

23:08agent that's doing all of these things.

23:10It's an agent that just says, "Okay, now

23:12I need to switch skills. I'm going to

23:14grab a different skill. So, first I'm

23:16going to research a topic. I'm going to

23:19grab my re- put my research hat on and

23:22do that. Then I'm going to interview a

23:23thought leader, get this context. Then

23:25I'm going to put on my SEO hat

23:27and do my SEO planning. Um I'm going to

23:31grab the voice profile skill and make

23:33sure I understand how Mike is supposed

23:35to sound or which company voice I'm

23:36supposed to use, and so forth. Same

23:39agent switching roles all the way

23:42through.

23:43Now, on our platform, uh this looks

23:46something like this on the back end. We

23:48call these recipes, and I'll show

23:52one of the first recipes. And this is a

23:53really common one that a lot of clients

23:55will use is how to automate publishing

23:58of your blog uh content.

24:01And so here, I can just say, "Hey agent,

24:04run the blog publish pipeline." And it

24:06will follow all of these steps. And it

24:09will pull together all of these skills

24:11and work through these I think it's 13

24:13steps. Takes about 10 minutes, and I can

24:16have a beautiful blog article that looks

24:17like I wrote it, with images, perfectly

24:20formatted, it all optimized for SEO.

24:24So, for a recipe here,

24:26when I call it by name, it will go

24:28through these steps.

24:30And you can see at various steps, it's

24:32like, "Well, load up this skill. Put on

24:34a different hat. Put on your SEO content

24:36hat."

24:37Load up your personal voice skill.

24:40Now you're going to put on your blog

24:42writing hat.

24:43Now you're going to get into quality

24:45assurance

24:47and review.

24:50And this is this is really important

24:51that you when you're working with

24:53agents, you should be able to tell them

24:55what success looks like.

24:58It's a very important thing. I think I

24:59have some slides dedicated to that in a

25:00second here.

25:02Where agents can get into these loops.

25:04This uh quality assurance loop is very

25:06very powerful concept in nails. So, you

25:08can see that in action here, where I'm

25:11saying that

25:12they have to assess their performance on

25:15voice alignment,

25:18they get a score from 0 to

25:20100. If the score is less than 90, it

25:22will go back and repeat over and over

25:24until it passes this stage.

25:27So, stages don't have to be linear, they

25:30can be iterative, and then when they

25:31pass, they can continue.

25:33That's what's happening here in this

25:35recipe. So, it's going to generate

25:36images, build the page, and do a whole

25:40bunch of stuff.

25:41And at the end, we get these beautiful

25:43blog articles.

25:46Uh if you go to our learning center,

25:48this article,

25:50"How a 4-minute interview becomes a

25:52published blog article." This is an

25:54article on the topic that I was just uh

25:57discussing, and it goes through. If you

25:58want to read more about how we did it

26:00and all the different steps,

26:02uh you can read that article.

26:04So, that's an example of a a of a one

26:07agent switching skills, and this is what

26:09most people think of as agents. But

26:12really, it's not quite there. I would

26:15say this is more like a

26:18AI workflow automation that people call

26:21it an agent, but it's

26:24maybe not.

26:25The the terms are still uh being loosely

26:27defined by the industry.

26:30Now, what are the limits of this

26:31approach? First of all is it's a bit

26:34slower. Right? When everything is

26:36linear, you have to wait for the

26:38research agent to complete their work

26:40before the SEO agent can do their work,

26:43and and so on. So, it it does slow you

26:45down. The second thing that you'll note

26:47with agents is

26:49they cost money. They're uh especially

26:51if you're paying through the API, they

26:53can be quite expensive, and token

26:55optimization is super super important,

26:56which is we'll have a dedicated workshop

26:58just to that. But when you're working

27:00with one agent that's switching hats,

27:02what happens is every time it switches

27:04hats and you interact with it, it's

27:06context is growing and growing and

27:08growing and growing. And if you were to

27:10track this

27:11uh on a a graph, your cost every single

27:14time that agent is having an

27:15interaction, the cost is slowly rising

27:18and rising and rising. You might start

27:205,000 tokens for the first session.

27:23Once you're 10, 15 back and forth in

27:26between those roles, you can be hitting

27:28200,000

27:30tokens for the exact same session.

27:33Another issue that we see is because

27:35these context windows get bloated, is

27:37sometimes the quality can deteriorate

27:40because it's like, "Who am I? Am I the

27:42SEO guy? Am I the CRO guy? Am I the

27:44designer?" Like, they start to get uh

27:48identity confusion at times.

27:50But,

27:52for simple processes, they they work

27:54they work well. And if you have a if

27:56you're on a Claude Max 20X account, you

27:59got tons of tokens, you're not worried.

28:01Some people just don't care about cost,

28:03um and if it's not something that's

28:05time-sensitive, this may not be an

28:07issue. So, we still use this single

28:10agent multiple hats in many scenarios

28:13that I'm not worried about, but it is

28:14something to be aware of. There's a much

28:16cheaper, faster, and higher quality way

28:18to do this, and that's where we get into

28:21uh multi-agent orchestration.

28:24So, this is all about having a

28:26coordinator and a bunch of different

28:27specialists that come in together and

28:29tackle a problem. And we'll go through

28:31sort of what that looks like.

28:33So, here's a visual

28:38uh

28:40There we go. Here's a visual

28:41illustration of what a multi-agent

28:44orchestration structure looks like.

28:47So, you might have the human here

28:49talking to project manager agent. The

28:51project management agent's

28:52responsibility is to orchestrate between

28:55a series of specialist agents.

28:58So, for example, say we wanted to build

29:00a website.

29:02I I'm going to need to do industry

29:03research, analyze competitors' website.

29:07If you've got an existing website, I

29:08would probably need to crawl crawl

29:09through it and do some analysis of how

29:13it's performing from a conversion rate

29:14perspective, from an SEO perspective,

29:17from a content perspective. And these

29:19can all be different agents with

29:21specialized roles.

29:22I might need to draft up some

29:23wireframes,

29:25actually design the pages, pass it to a

29:28developer agent, pass it to a quality

29:30assurance agent that gets into that

29:31iterative testing loop.

29:33And uh what else does it say down here?

29:36Agent roles. Okay, so

29:38Now, this is a very different structure,

29:40and this is truly agent or

29:42orchestration.

29:43And the way that that I'll show you sort

29:45of like a live demo here on our site.

29:47So,

29:49um the way I would pull that off, I

29:52would first grab a project manager.

29:58And the project manager's main

30:00responsibility is to communicate back

30:02and forth

30:03with other agents.

30:06It has access to

30:09delegate and follow up

30:11and organize tasks.

30:15So,

30:16I want to I want you to help me kick off

30:19a

30:21build of a new website. Actually, let's

30:24get a domain.

30:24>> [laughter]

30:25>> Somebody want a new website.

30:26Any volunteers? You know, what's the

30:27first domain to come up here?

30:31Let's build a website in 10 minutes.

30:35We have

30:36so many emails here. One second. Just

30:38one domain.

30:38>> go back up.

30:41Uh let's do luxdecor.com.

30:44luxdecor.com l u x d e c o r. .com

30:49>> u x.

30:51Um I'll I'll put it into the chat here.

30:54Did I say it right? d e c o r.com? Yes.

30:57Okay. You got it now. I should have

30:58[clears throat] recognized that one.

31:00Good to see you. Okay.

31:01Um

31:02and I'm not even going to edit this

31:04prompt. Agents can handle it. So,

31:09All right. So, I'm trying to I'm on a

31:10workshop right now and I want to

31:12illustrate a multi-agent orchestration

31:15structure here and I thought

31:17building a

31:19fake website right now would be a good

31:20illustration. So, um

31:23I

31:25want you to first spin up a researcher

31:28agent that will go out. Well, first

31:30actually analyze the luxdecor website.

31:33Understand everything you can about

31:34them. Come back at that information.

31:37Um I want you to spin up a second

31:39researcher agent that is analyzing the

31:41industry for luxdecor. Like as as soon

31:43as you know what their business is all

31:45about, go and analyze their industry,

31:47find their top three competitors and um

31:52look I I want you to build

31:54sort of like a market research report

31:56that says, these are all the top

31:57competitors. This is the luxdecor

31:59website. What are all the sort of

32:01features, benefits, and content that

32:03they have? What's different? And come up

32:05with a list of best practices and all

32:07the things that we should have on this

32:08new improved website. Uh then I want you

32:10to spin up a designer agent that will go

32:13to their website and create their brand

32:15identity a document and that should be

32:18accessible via public URL so we can see

32:20what this brand identity is. Then I want

32:22you to spin up a CRO agent that will

32:25analyze the website from a conversion

32:26rate optimization perspective. I want

32:28you to then spin up an SEO agent that

32:30will do the same thing. Once all of

32:32those reports are back, like I want you

32:34to check in with those agents every 5

32:36minutes or so until that is done. Once

32:38you've got all that context back from

32:39the industry research, from the audits,

32:42and um then I want you to take all that

32:44context and come up with a a set of

32:46wireframes with a design agent um and

32:49that design agent, when the wireframes

32:51are done, um go on and pass it to a

32:55well, the design agent can continue the

32:57next task would be to actually turn it

32:58into a full-blown beautiful website. Um

33:02you can put put it on my temporary sort

33:04of my AI one server here, make it

33:05publicly accessible.

33:08And uh then I want you to bring out the

33:09quality assurance agents once the site

33:11is designed.

33:13And

33:14analyze it, find ways you can improve

33:16it, get into a sort of like an iterative

33:18testing loop until the quality is

33:20amazing. And then when we pass that

33:22final test, I want you to kick it over

33:23to a developer agent who will actually

33:25code it, build it, make everything work.

33:28These are just sort of rough

33:28approximations of what I think you

33:30should do. If you can find better ways

33:31of doing it, go ahead and do it.

33:33Um go ahead. Now, Hey Mike, we we you

33:36need a hyphen between lux and decor.

33:40Okay.

33:41That is important information.

33:44All right. Now, that was a mouthful of a

33:46prompt, right?

33:48And a lot of you are probably thinking,

33:49well, I don't even know all the steps

33:50that I need to do to build a website.

33:52Well, that's fine cuz we can build an

33:53agent for that. I would be we call this

33:55a recipe. So, this is a playbook and if

33:59you have somebody who can teach you the

34:00playbook for a specific process like a

34:02we call them like the process owner or

34:04the human in the loop expert for web

34:06design in this case.

34:08Or you could just research it. We could

34:09go out with a research agent and say,

34:10what are all the best practices for

34:12building a website using agents? And

34:13it'll tell you all of those individual

34:15steps. Well, those steps could be baked

34:18into a recipe.

34:20And I call this recipe my software

34:21development pipeline.

34:24So, while that is going and hopefully

34:26we'll get a beautiful new website for

34:27Suna here in a minute. Well, probably 20

34:29minutes.

34:31Um

34:32So, here's an example of that very same

34:35recipe. Hopefully, it's on this I'm on

34:36the right server.

34:38Software development pipeline.

34:40So,

34:42if you can have an expert tell you

34:44what needs to happen

34:46in all the different stages of a

34:47process, whether it's anything like

34:50drafting a sales proposal or

34:53reconciling your accounting financials,

34:56you can turn those step-by-step

34:58instructions into a recipe. So, I could

35:01just come back there and say,

35:03kick off the software development

35:04pipeline. I just say those words and it

35:07will know what to do and it will follow

35:09step-by-step

35:11all those instructions and just guide me

35:12through it. So, you don't even have to

35:14know what to do. You just need to know

35:17which workflow or recipe to call.

35:20So, here like it knows I first need to

35:23gather requirements.

35:25I gathered the requirements in that

35:27example by

35:29doing a bunch of different audits.

35:31Research industry competitors website.

35:34Normally,

35:35you would have an interview with Suna,

35:37ask her 50 questions through

35:38requirements gathering process. Then it

35:40would pass it over to a technical

35:42scoping agent which would turn the what

35:44Suna wants here into the how we're going

35:46to build it.

35:47Then it's going to pass it along to the

35:49design and wireframing agents that are

35:50going to do that. Then it's going to

35:53break up if if the task is complex, it's

35:55going to break it apart into smaller

35:57tasks and organize it in Asana.

35:59If it's big enough, it's going to set up

36:01a GitHub repo for it to store the code.

36:04It's you know, all of these steps can be

36:06defined. And you can research them,

36:09right? You don't even I mean, it's good

36:11to have a human in the loop expert

36:12review all this stuff.

36:14But if I go back to that previous one, I

36:16could have just said, kick off the

36:18software development pipeline, walk me

36:20through step-by-step what I need to do

36:22to build an amazing website. And it will

36:23automatically get the information that

36:25it needs from you. It will automatically

36:27call the agents that it needs and

36:29accomplish the same thing without you

36:31needing to know how to create that crazy

36:34prompt that I just threw at it right

36:35here.

36:37So, that's an example of multi-agent

36:40orchestration. I think you can start to

36:42see why that's getting to be a bit more

36:44powerful than

36:46single agent

36:48workflows.

36:51So, that example project manager,

36:52researcher, designer, SEO, CRO, and many

36:55others. And remember, each of these

36:57agents, what were they? They're a

36:59collection of skills. They have access

37:01to different data systems.

37:03They have access to different tools.

37:05And they're rolled up into a role. So,

37:07the designer is really a function of 15

37:10different skills as is the SEO agent.

37:12And they all work together. They stay in

37:15their lane. And they hand off

37:17responsibilities back and forth through

37:19the project manager.

37:23So, why is more better here?

37:26When you take

37:29the context and you break it up through

37:31the project manager and you assign it to

37:33the individual specialist agents, their

37:36context window starts

37:38like open claw, it's not the best for

37:40tokens unless they fix that. They start

37:42about 12,000 tokens just to say hello

37:44world and it will keep rising and rising

37:46and rising. If you have that one agent

37:48that's switching hats, it'll keep going

37:50until it hits its limit. Around some of

37:52these agents are 256,000 tokens before

37:55they'll start compaction and um

37:58but if you take the responsibility for

38:01from the project manager and delegate

38:03that to eight different agents, each of

38:05those eight agents are starting at 5,000

38:07tokens on our platform and they will

38:09continue to rise over time. But if you

38:10add up all the token usage across your

38:13multi-agent orchestration,

38:15I mean,

38:17way way cheaper. I would venture to say

38:1810x cheaper in token cost um because of

38:21the the reduced context window.

38:24Another thing to consider is every time

38:25an agent has to connect to another

38:27system,

38:29there's overhead.

38:30Right? This was um we'll talk a bit this

38:32in the the next talk about token

38:34optimization, but if you for example are

38:36you're just having fun with Claude

38:37co-work, you're not yet into multi-agent

38:40orchestration, well, as your co-work

38:42boots up,

38:44just getting co-work started,

38:50every single connector that you have

38:53turned on here,

38:55I I'm pretty disciplined. I just just I

38:58keep most of them turned off if I'm just

38:59using Chrome. But let's say I had my

39:01Asana, my Gmail, my calendar, different

39:03connectors turned on for this is a

39:05pretty basic agent here.

39:08Well, each of those connectors is going

39:11to have to load up a bunch of

39:12information, make connections, bloat my

39:15token window.

39:17And if also if you have a generalist

39:21agent, it has to load up way more

39:23skills. Like say for example, the

39:25developer agent also needed to manage my

39:27DNS. Well, now it has to load up the AWS

39:30and the GoDaddy skills. And every skill

39:32that is attached to an agent

39:35is

39:36a big prompt that gets added into that

39:38context window.

39:39Right? So, think of that blog bot. It

39:41had a whole bunch of skills to boot up.

39:43There was probably 20 or 30 different

39:44skills to get through the whole 14-step

39:47process. Every skill, every connector

39:50adding to that context window. Whereas,

39:52when you've got the specialist agents,

39:53well, if you're just my DNS agent, and

39:56you're just responsible for GoDaddy

39:57subdomains and AWS, well, you've got

40:01very specific skills, a very defined

40:03prompt, and you start small, and then

40:05you build up. So, it's not just the

40:08breaking apart the tokens and starting

40:10smaller and building up, but it's like

40:11the initial tokens. Like, a a generalist

40:15agent will walk in and be like, "Hi, I'm

40:17Amy. I can do anything, and I'm going to

40:19cost you 200,000 tokens just to like say

40:21hello world." So, breaking that up um is

40:25is is big,

40:27especially when you start to scale these

40:28things and get into the hundreds or

40:30thousands of dollars a month in token

40:32cost.

40:33And that role specialization as well, it

40:35it it we've seen a big impact on

40:37quality. When I have a a testing agent

40:40whose only job is to test 10 different

40:44things and do it perfectly, when I have

40:46that nested within another agent, and I

40:49compare the output of the work, quite

40:50often, you'll see the the specialist

40:52agent just does a better job, especially

40:54if the the project or task gets complex.

40:57They'll start to forget. If the context

40:58window gets too big, their performance

41:00starts to deteriorate slightly. We've

41:03seen that over time. Now, the models are

41:05getting better and better. The smarter

41:07the models get, the less we have to

41:09worry about this specialization. So, it

41:11could be that if somebody's watching

41:14this video 3 months from now, they'll be

41:15like, "Oh, that doesn't matter anymore.

41:17Agents are just smarter." But, as of

41:19right now, that specialization still

41:21matters. And if you look at reports of a

41:23performance over time based on token

41:25window, typically, the the bigger the

41:27the context window gets Sorry, not token

41:29window. The context window gets bigger,

41:31then performance will slowly deteriorate

41:34over time, and then it sort of hits a

41:35wall.

41:39A big one is just

41:40speed, right?

41:43Hopefully, we'll see we'll see if we can

41:45get that

41:46example there for Sunak.

41:49But, if if you've done any project

41:50management, and you map things out in a

41:52Gantt chart, you can see that if if

41:55you're daisy chaining lines together,

41:56one line's going to be a lot longer.

41:58Multi-agent orchestration, we're going

42:00to have eight different lines, and some

42:01of them will have to wait. Like, we

42:03can't have the design agent kick things

42:05off until the research agents have done

42:07their work. But, I can spin up three

42:09different research agents in parallel

42:10that are doing

42:12the same things, and then they

42:14they work together, and then they come

42:15to the next line, and it just sort of

42:16condenses everything, and and goes a lot

42:18faster.

42:20And you can't really do an effective

42:22quality assurance loop when one agent is

42:25like, "I'm going to act as the

42:26developer, and I'm going to act as the

42:27person that's testing." And they go back

42:28and forth, back and forth. Beyond the

42:30context window shooting up is uh it's

42:32just not an an efficient uh path. This

42:34was the first multi-agent orchestration

42:36I did when I had two agents. One was

42:39building, one was testing, and one was

42:40orchestrating in the middle, and I just

42:42saw huge improvements in quality and

42:44speed uh with those testing loops. And

42:46these testing loops are critical for

42:48agents. Those of you that have first

42:50dabbled with agents, usually their first

42:52output kind of sucks. Like, you could go

42:54to Claude

42:55co-work, and you get it to build a sales

42:57proposal for you, and you create a scale

42:58and an agentic process for creating

43:00custom proposals. If you don't have a

43:03quality assurance testing loop in there,

43:04your first version is going to look like

43:05garbage. But, if you can if you measure

43:08what success looks like, say this is the

43:10template, this is the standard, did the

43:12output look as good as my expectations?

43:15If not, tell me all the things I need to

43:16improve. We create these iterative

43:18loops, and they get better. So, this is

43:20a

43:21important concept. You do need to

43:23understand and again, drinking from the

43:26firehose, write some of these notes down

43:27and dig a bit deeper offline there.

43:32And this is uh the example of of that

43:34loop. A draft, a different agent reviews

43:36it, compares it to success,

43:38says what's wrong, go ahead and make

43:40version two, and version two comes out,

43:43the agent comes back and tests again,

43:45and continues that loop until it passes

43:47various uh quality uh thresholds.

43:50Typically, the first draft in an agent

43:52uh is not at the quality standards that

43:54you'll need.

43:56Okay, but you do need to have clearly

43:58defined success criteria for these

44:01things to work,

44:02uh so that they can evaluate themselves

44:03and have those iterative improvements.

44:07Okay, so,

44:08where do we go from here?

44:11Well, I mean,

44:13AI generated all these slides for me

44:14half an hour before I I I don't actually

44:16agree with thinking big first. I would

44:18say think small.

44:20Um start with something like Claude

44:23co-work. That's a really safe place.

44:26Um

44:27A lot of entrepreneurs I know have been

44:28dabbling with open claw and Mac minis

44:32uh or the Her- Hermes agent. These are

44:35open source.

44:37Um but uh what I'd warn you,

44:39oh, I already fixed that one. So, yeah,

44:41okay, start with one. Don't think too

44:42big. Start with one. Start with

44:44something simple. Claude co-work is

44:45probably the safest place uh for anyone

44:47to start if you're not super technical.

44:51And build one simple agent. Get it to

44:53work, and

44:55you don't need to graduate to

44:57multi-agent orchestration right away.

44:59You know, like I started with open claw

45:01and Claude co-work and Claude code

45:03individual agents, and at some point,

45:04you'll be like, especially those of you

45:06on

45:07Claude co-work,

45:08you'll get your agent crushing things,

45:11and you're going to want more.

45:12Right? And like, I had two computers

45:15going, two Claude co-work sessions, and

45:16I just wanted more.

45:18When you're at that point, you're like,

45:19I I need more agents,

45:21then move to multi-agent orchestration.

45:23You don't need to go there right away.

45:24You will have amazing experience just

45:26building your first agent or a couple of

45:28individuals. Start one and scale up.

45:31But, there's an important decision that

45:32you all need to make around like, where

45:34do I build? Now,

45:37selfishly, you know, I have my biases.

45:40We've built a platform that does this,

45:43but

45:44honestly, it's not for everyone, right?

45:47Like, if you are a technical

45:49entrepreneur that loves getting into the

45:51weeds of systems and processes, but

45:53maybe you don't have any experience with

45:54software development, but you'd love to

45:56learn how to do these things, the

45:58learning curve is getting shorter and

46:00shorter.

46:01I've seen multiple friends of mine, EO

46:03friends that

46:05just went off the deep end, 16 hours a

46:07day for a month. They came back, they're

46:08like, "I'm

46:10building software. I'm building agents."

46:12These are

46:13entrepreneurs with zero software

46:14development background, but they're

46:16technical people. These are the sort of

46:18people that

46:19just love diving into something and

46:22figuring it out, and want to go deep

46:23into those weeds. You can do it on your

46:25own. And if you do want to go down your

46:27path, there's some very low-cost

46:29open-source ways to do that. But, if

46:32you're not diligent, you can create

46:34massive security holes

46:36and vulnerabilities for your business.

46:38If you've been tracking the news,

46:39especially around some of these new AI

46:41models like Claude Mythos that are

46:42coming out, AI is getting better and

46:44better at hacking as well, and we have

46:46to be very careful with security.

46:48Security's probably the number one

46:50consideration we have with our agents.

46:52So, if you go down the path of say

46:54building something like a Hermes agent

46:56on a Mac mini or a Mac Studio,

46:58incredible learning experience, but if

47:01you don't have

47:03a good solid understanding of

47:05how to make it safe,

47:08you need to build it in a way so one,

47:09it's got to be on a separate computer,

47:11um does not have access to your email or

47:14your two-factor authentication flow,

47:16whatever that is,

47:18and start really small. Use it as a

47:20learning experience, but at some point,

47:21you're going to have to bring in

47:22security experts to make sure your

47:25hardware is locked down, that you're

47:27scanning all the files that come in and

47:28out to look for prompt injection.

47:31There's a lot to to learn there.

47:33Whereas, something like Claude co-work

47:34out of the box has built-in uh security

47:36protections. It's a lot safer than than

47:39that. So, I usually tell people,

47:42it's fine. Even if you're non-technical,

47:44and even if you have no security

47:45background, you can play with these

47:46open-source agent platforms.

47:51hurt you, cuz you will probably have

47:53major security vulnerabilities. So,

47:56treat it as a bit of a testing ground.

47:58And and then from there, I say like, if

48:01you are

48:03if you don't have a strong in-house

48:05technical person,

48:07I mean, selfishly, I'd say just hire an

48:09agency like ours, and we'll run it for

48:11you. But, in reality, every business

48:13should have in in in this age of agents,

48:16you need like one person that's strong

48:18in-house that's managing your agent

48:20team. So,

48:21um

48:23it's kind of choosing like, are we going

48:24to build the talent from within? They're

48:26very hard to find these people right

48:27now. There's so much demand for them,

48:29and these are new professions, people

48:31that manage agent teams for you. But,

48:33building that expertise in-house, I

48:35think, is going to be a core competency

48:37of businesses uh in the next year. So,

48:42so, you kind of have like, build it on

48:43your own as a founder, get into the

48:45weeds, figure it out, but be really

48:46careful with security.

48:48Hire somebody or train somebody up

48:50in-house over two, three months. They

48:52can do it if they're the right person,

48:53and it's not you.

48:54Or

48:56just get somebody else to build it for

48:57you. But, even if you get somebody else

48:59to to build and run these for you, I

49:01think it's really important that you

49:02understand at least what's going on. So,

49:03we always tell clients like, we ask them

49:06when we're setting them up, "Do you want

49:08to be involved with the learning

49:09process?" Most of them say yes. And I I

49:12think like,

49:14because

49:16I don't know if if if what you saw me do

49:18there with some of those agents looked

49:20complicated. I mean, the the pieces

49:21building up to it is fairly complicated,

49:24but once you've got your agents, you

49:26just talk to them like employees. Go do

49:27this, go do that. And once you know how

49:29that all works, you can make them better

49:31very easily. You can say, "I want you to

49:33improve yourself by adding the following

49:35three features. Let me know what you

49:36think and boom, you just upgraded an an

49:38agent. It's not rocket science. So, I do

49:41think that founders need to understand

49:43at least that sort of 20 30% of what's

49:45going on so that they can properly

49:47coordinate with specialist teams or

49:49other agencies.

49:51So, again, this is this is a marathon.

49:53The the learning should not stop. Um

49:56those of you that have been with me for

49:57the last four or five years, you'll know

49:58these workshops just keep going. Up

50:00next, uh May 21st, we'll be going deeper

50:03into how you can optimize your agents to

50:05reduce the cost. I've seen uh results

50:08where people are spending 90 or they're

50:12getting a 90 plus percent cost reduction

50:15of their agents once they apply these

50:17token optimization strategies. We're

50:19also going to be doing a panel on AI

50:21security. So, uh Rishi Khan is a fellow

50:24EO member from Philly. He works for the

50:28NSA and uh he's one of the most

50:31experienced security people in the field

50:33of AI. He's going to be on a panel and

50:35we're looking for one more guest to join

50:36me and we'll be going deeper into how to

50:37make sure that you don't uh screw up

50:40your agents and

50:41get yourself into some sort of hack

50:44situation, which is big. So, it's an

50:47incredible journey.

50:48I recommend that you do not do this

50:50journey on your own. Do it in teams. I

50:53learned so much from talking to groups

50:56of people. I'm in multiple different AI

50:58learning forums. This QR code here is

51:00our little community. Uh this one is for

51:03both EO and non-EO members. If you scan

51:05that QR code if you're not already part

51:06of that community, that's where you just

51:08post your questions. Everyone's learning

51:09and sharing together.

51:11This is a long journey. This is a

51:13marathon. It is not a sprint. You got to

51:15do it in steps and stages and we learn

51:17so much faster together as a team. So, I

51:19hope to see many of you on the next

51:21couple of workshops. We've got two next

51:23month. And now we're going to open

51:25things up for Q&A. Hopefully, I haven't

51:28thoroughly overwhelmed you too much.

51:30Did you show the example of the website

51:32and stuff?

51:34Oh, right. Yeah, let's go back and let's

51:36see. Good call.

51:37Um

51:40Now, that was a fairly long prompt. A

51:43lot of it was just showing you how much

51:45you can do in one thing. Let's check in.

51:48All right. So, this project manager

51:50agent is now on phase three.

51:53It you can see in the background it's

51:54been doing a bunch of stuff.

51:58Um

51:59First, it was checking memory to see if

52:01we had anything in the brain already on

52:03Lux.

52:05Reading through its different skills.

52:09Spinning up all five phase one agents in

52:12parallel.

52:13That's what it did here at this stage.

52:16These are all condensed uh

52:18you don't need to see the code behind

52:19that.

52:21Agent one, you are a brand researcher.

52:23Agent two, you are the industry

52:25researcher. This one, you are the brand

52:27identity uh designer.

52:30It's telling them what they are, what

52:31their roles are, pulling up their

52:32skills, booting up the agents.

52:35There, it's loading the CRO SEO agent.

52:38So, a bunch of things happening in

52:39parallel.

52:40SEO is done.

52:43Phase one two done.

52:46CRO report done.

52:50Okay.

52:51Phase one complete. All six reports are

52:53in.

52:54Um this is all of its thinking behind

52:56the scenes. Let's just see some of these

52:58reports that it was

52:59building. This one is

53:02the first and remember, first version

53:04always sucks. So,

53:05judge my agent here cuz we can make this

53:07better. So, it came up with a brand

53:09identity document.

53:11I don't know, Sana, if this is aligned

53:13with

53:15your brand.

53:17But, brand identity from a sort of text,

53:19voice

53:21um

53:23So, it built this.

53:25And that's a core You can't build a

53:26website without uh

53:29a style guide.

53:30So, we created that asset.

53:32How are we doing, Sana?

53:34Do you like your new brand identity?

53:37That's really cool.

53:39I love it. I can't believe you did that

53:41in a few minutes. It's amazing.

53:42>> Yeah, so so that little piece, yeah,

53:43there's probably 40 different things

53:45that happened. Okay, so we've got

53:47the brand identity.

53:49Uh it also needed to create the industry

53:51research.

53:53Um

53:58Oh, it just finished the wireframes.

54:00Can you give me a list of URLs of all

54:02the different assets that you created uh

54:04for your research like the industry

54:05research, the brand identity, the CRO,

54:07the SEO? Uh

54:09give me all those URLs so I can uh share

54:12publicly.

54:14Uh this agent is still running, so

54:16um

54:18I'm going to queue that up.

54:21You know what? I'm just going to stop

54:22it.

54:24Insert that.

54:25Get that done. Let's see how it did on

54:26your wireframes.

54:30Oh.

54:32It's not loaded here.

54:37So, now it's handing the wireframes off

54:39to the visual designer next.

54:41Um

54:42You should be able to see some more

54:43links here. Oh, here's

54:48Oh, it created alternate brand

54:49identities.

55:00Yeah, this guy needs a little bit more

55:01time. Still working. Um

55:03but

55:04you'll see the same way that we got the

55:07brand identity

55:10document,

55:12there'll be a bunch of beautiful

55:13documents like an SEO report, a CRO

55:15report, industry research report. It

55:16rolls all of that up. Um yeah, we'll

55:19show you we'll show you that those

55:21reports in a second. Looks like they're

55:22all sort of hidden behind the scenes cuz

55:24the final output will be the design.

55:26That's what it's still working on. But,

55:27you can see how it's working through

55:28those steps. And that's really like when

55:31you start prompting properly with

55:33agents,

55:35you're prompting in a very different

55:36way. Like in the past, it's like one

55:38shot. Go do this. Go do that. And you

55:39can see here like you can give it

55:41something that it can work away. I've

55:42had some agent process or working away

55:43for 12 hours. I come back in the morning

55:45and they've like built a new piece of

55:48software.

55:49So, yeah, we'll get you some more

55:51examples there as it's working away. Uh

55:53Gabe, good to see you.

55:54Or

55:55almost see you.

55:57Got a question.

55:58Hey, Mike.

56:00Uh great workshop as always. Sorry, I'm

56:01just driving right now so I can't turn

56:02on the video, but um

56:05the question for you. I I recently read

56:06a study that said that the multi-agent

56:10orchestration

56:11is almost a better is better through

56:14self-determination in terms of

56:16the agents decide

56:18what the best sub-agents are to use for

56:21a particular task.

56:23And

56:24they can do that more efficiently than a

56:25human could in terms of deciding which

56:27agents to use.

56:29And so,

56:30I would love to hear your take on when

56:32you kind of provide the guardrails of

56:34saying, "Okay, I want you to create this

56:35agent, that agent, that agent."

56:37versus just saying, "Here, I want this

56:39outcome with this skill.

56:41You decide which agents to use." Like,

56:43have you tried that approach and do you

56:44have any kind of thoughts around

56:46efficiency, quality, outcomes? Like,

56:48what are you seeing

56:50uh producing the best outcome?

56:52And and why do you choose the

56:53multi-agent approach as opposed to

56:56letting it just decide on its own?

57:00So,

57:03I think if you if you

57:05don't predefine

57:09the the the the steps like for example,

57:11the software development process. Uh

57:12we've gone through 20 30 different

57:14iterations of that in testing and to see

57:16what works best. And I like if if you

57:19have a true human in the loop expert

57:21that is defining those steps and stages,

57:23I think you can get a better, more

57:25consistent outcome by defining by

57:27defining them. I think if if it's like

57:30somebody who doesn't know exactly what

57:31they're doing, you'll you'll immediately

57:33get better results. If you stop like

57:35don't you know, it's like when clients

57:37try to tell agencies what to do, the

57:39tail wagging the dog sort of thing. Like

57:40if you're not the expert, get out of the

57:42way and let the expert determine

57:46the way that it came up with a software

57:48development pipeline, which was the sort

57:50of the the guardrails for how we would

57:52go through all those different steps,

57:54that was through just tons of research

57:56and testing. And so, we and and by

58:00finding those best practices and seeing

58:01the different outputs, we can cement

58:03that so that every time we get a

58:04consistent output. But, yeah, like as

58:07the models get smarter and smarter, the

58:09amount of work that we have to do to get

58:10them to do a good job is getting less

58:12and less. I said a few years ago, I

58:14think prompt engineering will eventually

58:16go away on its own because agents will

58:18just be smart enough to to ask the

58:20questions that they need and to to just

58:22know what to do as they get to know you

58:24better as your company brain and

58:25knowledge grows. Yeah, so I think that

58:28over time, yes, we're going to see more

58:30of that where the agents just

58:31intuitively know things. I think if

58:33you're not a human in the loop expert um

58:36with say web design, you're going to be

58:38way better off having the agent decide

58:40the path for you. But, if you want to

58:43get something that is repeatable and

58:45very consistent like a sales deck that

58:47comes out the same every time, um if I

58:50left that those decisions up to the

58:51agent, I might get a beautiful sales

58:53deck that's slightly different every

58:54time. So, I think there's pros and cons,

58:56but it yeah, it's changing. Um as agents

58:59get smarter, we

59:01have to give them less and less

59:02direction. Less I mean, before December,

59:05we were doing things like workflow

59:07automation using tools like N8N,

59:09make.com, all of that. That stuff is

59:10pretty much obsolete. Uh well, not fully

59:13obsolete, but it's crashed. You know,

59:15that whole market is

59:16dying.

59:17Um and that was about really, really

59:19specific guard guardrails that would

59:21guide your agents through every little

59:22individual step. And now with these

59:24agents as a broad concept, like I didn't

59:26say exactly how to build the brand

59:28identity for Lux. I just said, "Go

59:30figure it out and you have the right

59:32skills. And it just did it. It wasn't

59:34even

59:35the It was loosely defined scaffolding

59:37compared to just 3 months ago. So, yeah,

59:39the general trend is what you have said.

59:42I would agree with that, but I think

59:43it's more about consistency. And if you

59:45have an expert system that you want to

59:46replicate consistently, I would turn

59:48that into a recipe just for that

59:50consistent outcome.

59:52But

59:52>> So, could you could you build those

59:53guardrails with

59:56a skill as opposed to many different

59:59agents or as you say a recipe? Like does

1:00:02it have to be

1:00:03many [clears throat] agents to produce

1:00:05that predictable outcome in your

1:00:06experience?

1:00:07No, and that's where we get into a

1:00:09single agent workflow, right? Like the

1:00:11blog, that's just one agent that's

1:00:14grabbing all the different skills that

1:00:15it needs at different places. But the

1:00:17disadvantage with that is your your your

1:00:19your token window is expanding up to the

1:00:21maximum pretty fast in that scenario.

1:00:23So, it does work. Yeah.

1:00:26Right? I could say

1:00:27>> based on context, Rob. Like basically,

1:00:30as soon as you get over a certain amount

1:00:31of tokens,

1:00:32that's where you have to build in the

1:00:34multi-orchestration piece. Well, you

1:00:36don't have to. I mean I mean with just

1:00:37within Claude CoWork, for example, I

1:00:39don't even think you really have that

1:00:41ability. It's just loading up different

1:00:42skills. Unless you're loading a whole

1:00:44bunch of separate Claude CoWork sessions

1:00:46cuz you happen to have 64 GB of RAM,

1:00:48props to you, but most people they're

1:00:51they can only run one Claude CoWork

1:00:52session at a time before their computer

1:00:54melts. And that that agent will just be

1:00:56switching hats. Like now I'm going to be

1:00:58the quality assurance skill, and I'm

1:00:59going to go back to developer skill. Now

1:01:00I'm going to do a design skill. It's the

1:01:02same agent switching hats, and that does

1:01:04work. And it can figure that stuff out

1:01:07on its own, but if you've analyzed the

1:01:09token cost of that versus that breaking

1:01:11it down into those specialists, it's

1:01:13like a 10x difference.

1:01:16Interesting. Okay, thanks a lot.

1:01:17>> Yeah. Thanks, Mike. Okay, and we just

1:01:18got those I asked for those example

1:01:21assets that we were creating for Suna.

1:01:23Let's take a look at those. Some of

1:01:24those just came back now. So,

1:01:26here were all the different

1:01:27>> The security audit, Mike. Also, the

1:01:28security audit. Someone requested to see

1:01:30that, too.

1:01:31>> Oh, right, right. Okay, let's quickly

1:01:32see our security audit. Who was that

1:01:34for? Fence More security review.

1:01:36Um

1:01:37Okay, there's their security review that

1:01:39came back.

1:01:41Um

1:01:4345 out of 100. Got some security issues

1:01:46to resolve.

1:01:47Now, we can send this report out. If you

1:01:49email or give Suna your email address,

1:01:52we'll send you this report.

1:01:54Yeah, just with asking a couple of

1:01:55questions, it goes through and builds

1:01:58this interactive website

1:02:00um

1:02:02just with a couple of words. So,

1:02:03complexity rolls up into simplicity. I

1:02:06could even just paste in a URL, and it

1:02:08will get me a report like this very

1:02:10quickly. And the magic isn't so much in

1:02:12these audit reports. The magic is when

1:02:14they're like, "Okay, this is crazy, but

1:02:15now what do I do?" Well, tell your

1:02:18developer agent to go and fix all these

1:02:19security issues. Do it on your

1:02:20development server. And when that's

1:02:22done, go back and run the security audit

1:02:24again. Put the comparison side by side,

1:02:26and suddenly you're like, "10 minutes

1:02:29went by. My security score went from a

1:02:3045 to a 97."

1:02:33And you didn't hire anybody. You just

1:02:35hired an agent. So, that the power is

1:02:37the implementation, not just the

1:02:38auditing.

1:02:39Now, let's take a quick look here at um

1:02:42some of the other assets that were

1:02:44created for the website.

1:02:47So, we did

1:02:49Looks it went off on its own and created

1:02:52two versions of the brand identity. Uh

1:02:55so, I guess we can choose between one

1:02:57and two. It did brand research on who

1:03:01you are, what you stand for.

1:03:04This is all context

1:03:06to help it do a good job.

1:03:08Target customer, value proposition,

1:03:10brand voice and tone,

1:03:12current visual style, site structure,

1:03:15content inventory. So, just doing more

1:03:17more research on you.

1:03:19We did industry and competitor analysis.

1:03:26Market

1:03:27Look at your top three competitors,

1:03:28McGee and Co.

1:03:31Jenny Kane.

1:03:34Uh pros and cons. I asked for that

1:03:36feature comparison matrix of what you

1:03:38have and what they have to start to

1:03:40think of what a better version of your

1:03:42site would could include.

1:03:44So, we can see, for example,

1:03:49uh they all have mobile-first sites.

1:03:51Your site apparently is uh a subdomain.

1:03:54That's not a best practice for Google

1:03:56anymore for SEO. Unified domain services

1:03:58plus shop. Um you've got that split. Um

1:04:04And this is non-visual, right? I could

1:04:06easily come back and say to it like this

1:04:09industry and competitor report, like um

1:04:13I want you to show you the power of our

1:04:14Nano Banana skills. Can you go back and

1:04:16and make the industry and competitor

1:04:18analysis report more visual by um adding

1:04:22custom images from Nano Banana like the

1:04:25image gen model from from Gemini

1:04:27throughout that aligns with the brand

1:04:29identity that really takes all of the

1:04:31text and brings it to life visually.

1:04:34Please make that upgrade to that report.

1:04:37And I could also just build that into

1:04:38the skill in advance so that all

1:04:40research reports will be more visual,

1:04:41and give it 3 minutes, and this thing

1:04:43will look visually stunning.

1:04:45Uh we did the CRO

1:04:48audit,

1:04:51finding all the various issues.

1:05:01So, again, I This is just getting all

1:05:02the context that it needs. This is the

1:05:04the requirements gathering, the context

1:05:06gathering. CRO, it did an SEO audit,

1:05:12finding all the technical issues and

1:05:13opportunities.

1:05:16These are all the building blocks. Then

1:05:18we got to

1:05:21Oh, it's already starting

1:05:23the wireframes.

1:05:29Oh. Now, it created a

1:05:33website rebuild workshop.

1:05:35Oh my goodness, look what it did. It it

1:05:38built the creative brief.

1:05:40It's These are all different sections.

1:05:41The CRO launch specs,

1:05:44the brand identities, different versions

1:05:46that you can choose from, brand identity

1:05:48summary.

1:05:49Um for the agent,

1:05:52more brand research, industry competitor

1:05:54analysis. These are all the audits that

1:05:55it complete Okay, so So, it created a

1:05:57book to show all the research that it

1:05:58did before

1:05:59the website project. Um

1:06:04And

1:06:06amazing.

1:06:07Okay, we're still waiting for uh

1:06:10The The final output here should be a

1:06:11beautiful-looking website. Uh continue

1:06:13with your work.

1:06:15Hopefully, I didn't confuse it, but it

1:06:16looks like it's still building from

1:06:17there. But you can see This is an

1:06:20important thing, too, with AI in

1:06:21general, is planning is more important

1:06:23than than building. The building happens

1:06:25really fast. So, what I see a lot of

1:06:26people do, and I see Brennan with his

1:06:28hand up, and I saw he did it when he

1:06:29first got access to these orchestration

1:06:32tools, he just dove in and just started

1:06:34building stuff really, really fast. And

1:06:36you can see here before I'm going to

1:06:38even lift a finger and start designing

1:06:40things, look at all the research and

1:06:41planning that went into it. Planning is

1:06:43is is really important step a lot of

1:06:45people skip over here with with agents.

1:06:47All right, let's continue with some

1:06:48questions. Uh Brennan, over to you. Good

1:06:50to see you.

1:06:54Thanks, Mike. Yeah, that great another

1:06:56great uh workshop. So, yeah, my question

1:06:58is, you know, so much of this can be

1:07:00done, quote unquote, in in um

1:07:03uh Claude and

1:07:06uh even the designing of these um

1:07:08>> [clears throat]

1:07:09>> uh multi-step workflows. It's exciting,

1:07:12and you you start you get you get down

1:07:14the path, everything's possible, and

1:07:15then you hit a complexity wall. That's

1:07:17what's been happening with me is when

1:07:19it's time for designing

1:07:21uh uh

1:07:22the the workflow, it's very easy and

1:07:24exciting. And then when uh it's time for

1:07:26action, I have, you know, complexity

1:07:28comes in, I have all these connections

1:07:29to make, um all these decisions to make,

1:07:32and so forth. And so then, the logical

1:07:34step would be, "Okay, let's work with a

1:07:36platform like AI what?" Because now I

1:07:39can enroll agents and so forth. And I

1:07:41guess the question is, as I said in the

1:07:43in the chat, is like, is this a trap for

1:07:46small businesses? Can

1:07:48you know, how much time and effort do

1:07:49you put in before you say, "Hey, I can

1:07:51This is This is consuming all my time

1:07:53and effort. I'm not concentrating on my

1:07:55job. I'm staying up all night." Um and

1:07:58and if you get in this position, what do

1:08:00you like can you get help? And what

1:08:03where you know, you've you've been in

1:08:05the space for a long time. Where do

1:08:06people go to get help for this uh this

1:08:08you know, this You have all these

1:08:10aspirations, you want to do all these

1:08:11things,

1:08:13um and it takes time and effort and

1:08:14knowledge, and uh so where do you go?

1:08:16And you have to move fast. Like this is

1:08:19Those of you that saw my first talks 4

1:08:21years ago, I was telling you back then

1:08:23you got to move fast, but now it's like

1:08:25we have months. We have weeks. Like

1:08:27people are moving so quickly here.

1:08:31Um

1:08:33Like what

1:08:34Yeah,

1:08:35like if you just hire an agency like

1:08:37ours, of course, I'm happy. That's great

1:08:39for me, but in reality, like I think

1:08:41what you

1:08:43you need to master these skills as an

1:08:45entrepreneur. It's a whole new set of

1:08:46skills that we have to learn. And if you

1:08:48just say pass it off, like I see Tomas

1:08:50here on the call. Tomas is an

1:08:51exceptional AI engineer. You can hire

1:08:54someone like Tomas, and he can help you.

1:08:56But if you don't take the time to master

1:08:59these skills, I think you're going to

1:09:00miss out on probably the biggest

1:09:01entrepreneurial opportunity of our

1:09:03lifetimes. This stuff is not that hard.

1:09:05You do have to dig into it. You know, a

1:09:08couple of months of focused learning,

1:09:09and you can come out here and sort of

1:09:11You know the Pareto principle there? You

1:09:13can probably get to 80% understanding

1:09:15here with 20% the effort. You don't have

1:09:17to like do the 16-hour days stuff that

1:09:19I've been doing to be a true master.

1:09:21Um and then I think like having somebody

1:09:24like an expert like a human-in-the-loop

1:09:26expert that can review before you take

1:09:28certain steps. Like for example, if you

1:09:30were building

1:09:32with your own website, well, you should

1:09:34probably have a designer friend who's

1:09:36going to sign off on your design before

1:09:37you say that are you qualified to say

1:09:40that that was a good design and that

1:09:41your customers are going to like that?

1:09:43You should probably have an SEO expert

1:09:45sign off on the SEO plan. You can't

1:09:46trust the agents perfectly. You need to

1:09:48have these human in the loop experts.

1:09:50So, I think that's where like an agency

1:09:52model comes in good cuz the agency has

1:09:54all the human in the loop experts built

1:09:55into it. But if you're going to build

1:09:57that on your own, you can. You just need

1:09:59to dive down that rabbit hole and you

1:10:01need to have experts that you can call

1:10:03for very thin slices of things. Like for

1:10:06example, I'm calling Tomas on strategy

1:10:08advice on when we build complex plans

1:10:11for building software. I'll say Tomas,

1:10:13can I borrow you for an hour to review

1:10:15this plan, make sure it makes sense

1:10:17before the agents take over. Right? So,

1:10:19I don't you don't have to even

1:10:20necessarily hire an agency full-time.

1:10:23You could just hire an expert. But you

1:10:25do need the right humans to support you

1:10:27cuz I

1:10:28sure you've all seen it. AI will

1:10:30occasionally do brilliant brilliant

1:10:32things and then make incredible

1:10:33blunders. And if you're not there to

1:10:34catch it, it'll screw up. So, yeah, I

1:10:38you can do it on your own. You need to

1:10:39have experts around you to support you.

1:10:41And it's an incredible journey that I do

1:10:43think most entrepreneurs should engage

1:10:46on.

1:10:47Um Well, is it Can you just can can you

1:10:49learn at a at a at a vision level and

1:10:52then work with technical technical

1:10:54people for the execution side? Like is

1:10:56that is that sufficient to understand

1:10:58what's possible?

1:10:59>> yeah. And that that model now a lot of

1:11:01clients are are a lot of um

1:11:03businesses are are hiring you're hiring

1:11:06agents as a service. The the the acronym

1:11:09I think is lower case A capital G A A S.

1:11:12Agass or ass.

1:11:14Agents as a service. So, you can hire

1:11:18ass

1:11:19and just understand it from a high level

1:11:21and get somebody to build it for you.

1:11:23And then but you do need to understand

1:11:25how to take an agent and make it better.

1:11:27How to improve it. You know, so even if

1:11:29someone's building all your agents for

1:11:31you and they're out of the box with

1:11:32human in the loop experts. Like that's

1:11:34what we do. You still need to understand

1:11:36it all. It's like it's imagine you hired

1:11:38somebody to build a website for you but

1:11:40you didn't even know what websites were.

1:11:42Are you going to get a good website at

1:11:44the end?

1:11:45Fingers crossed you don't get jacked,

1:11:48right? Pay way too much or have

1:11:50something that doesn't you didn't even

1:11:52need.

1:11:53Um so, yeah, I would go 20/80 at

1:11:56minimum. Learn the 80% with 20% the

1:11:58effort, master it. If you are going to

1:12:00hire an agency to do it for you, you

1:12:02need to know how to to talk to them, how

1:12:04to engage with your agents, how to how

1:12:06to get more from them. So, I I don't I

1:12:08don't think this is like

1:12:12Social media when it came along, did we

1:12:14all have to learn that? Sort of. It

1:12:16wasn't do or die.

1:12:18SEO, we all kind of had to learn a

1:12:19little bit about that about websites in

1:12:21the past, the internet, typing. Some of

1:12:23these things were

1:12:25you could get away with not being a

1:12:27master at but I think the the

1:12:29You've heard me say this before. Like

1:12:30the advantage you get from mastering

1:12:32agents can be like a 10x advantage over

1:12:35your competitors. That if you don't nail

1:12:37that as an entrepreneur, I think you're

1:12:39going to be left behind. I I We're going

1:12:41to see winner take most scenarios in a

1:12:43lot of industries. And the ones that are

1:12:45taking most, it's where the CEO put on

1:12:47his hat, said, "Honey, can I get a hall

1:12:50pass for 2 months? I'm going deep into

1:12:52agent world."

1:12:52>> a call I didn't realize they were here.

1:12:55And uh so, yeah.

1:12:58Thanks. Thanks, Mike. Yeah. Appreciate

1:13:00it. Uh Tina, hello.

1:13:04Hi. Um

1:13:06wonderful wonderful um

1:13:09workshop. Thank you for this. I am new

1:13:13into this world and am just starting to

1:13:17um understand that I need several

1:13:21different agents.

1:13:22Um are there

1:13:24best-in-class resources or places to

1:13:28um understand

1:13:30like you showed us uh and I'm probably

1:13:32not going to use the right language but

1:13:34there was a framework that you showed

1:13:36that had, you know, all the different um

1:13:39agents that you're working.

1:13:40Um if I have very repeatable processes,

1:13:43is there a best-in-class place that I

1:13:45can look to see Hey, what what's a

1:13:49decent framework to start with? Like I'm

1:13:52I'm realizing I need these things but I

1:13:53don't know all of the things that I

1:13:55might need. And then my my next question

1:13:57is um

1:14:01are there best-in-class resources

1:14:03um that I can go to at least get some

1:14:06out of the box skills to start uh

1:14:09loading up the agents within? And and

1:14:12then obviously would look to customize.

1:14:17Excellent. I think that the best place

1:14:19for people to get started that are not

1:14:20super technical would be with Claude

1:14:22CoWork. That's the most common starting

1:14:24point that I'm seeing right now.

1:14:26And

1:14:29yeah, if you're just downloading Claude

1:14:32desktop, it works for both Mac and

1:14:34Windows PC now as well. It doesn't work

1:14:36so well in the home edition of PC. It's

1:14:38really good for Mac or Windows Pro 12.

1:14:41I'm not a Windows guy but there's

1:14:43certain versions of Windows it sucks

1:14:44sucks with. And where do you get those

1:14:46sort of initial skills from? I mean,

1:14:47CoWork teaches you about skills. I

1:14:50Janet, we can actually send out the

1:14:51training course for CoWork. We've got um

1:14:53on our website here, we've got sort of a

1:14:55CoWork 101 guide and uh training program

1:14:58that can guide you along here.

1:15:02I just had to update cuz they launched

1:15:04new

1:15:05features. So, I just had to update it

1:15:07like 2 days ago.

1:15:08Um this one here like the CoWork 101

1:15:10getting started guide.

1:15:13This is a good one. Like if you get uh

1:15:15into CoWork, follow this guide.

1:15:18Um this is all about agents and the

1:15:20difference between chat and agents and

1:15:22how to prompt and all that. This is a

1:15:25really good starting guide and Claude

1:15:26CoWork is a good base. And then like

1:15:28when you go to

1:15:30Like this is where it all starts is is

1:15:32the the skills. Like

1:15:34They have a whole skill store. Like

1:15:36there's tons and tons of skills that you

1:15:37can just get open source off the web.

1:15:39Skills themselves are really easy. And I

1:15:41also just I typically just build my own

1:15:44skills. You can say something like I

1:15:46want you to develop a skill for an SEO

1:15:49agent. Research the web, tell me best

1:15:51practices and develop the skill and put

1:15:54it in research mode.

1:15:58Where is research mode on this?

1:16:00Does CoWork not have research mode?

1:16:03They keep changing the layout.

1:16:04Everything's moving every day.

1:16:07Well,

1:16:08maybe that's on chat we do deep

1:16:09research. But yeah, I I a lot of my

1:16:11skills, I just start out by doing deep

1:16:13research on that final industry best

1:16:15practices. You can go into a CoWork

1:16:18under um

1:16:22Oh, yeah, sorry. It's under the

1:16:22customize section. Click on skills.

1:16:25Skills. Um you can search through their

1:16:28library of existing skills or you or you

1:16:31can just create your own. Like here I'm

1:16:33going to browse skills and

1:16:36type in

1:16:38I don't know.

1:16:39SEO.

1:16:40There's nothing there. Oh, but you can

1:16:41sort of see um

1:16:45from different partners.

1:16:47Uh

1:16:49Where's their full skills directory?

1:16:51Yeah, that's why I'm asking because I've

1:16:53been in there and I haven't really found

1:16:55a rich repository there and maybe I'm

1:16:59just doing it wrong. No. Oh, it's the

1:17:02plugins now. They've moved it, right?

1:17:03So, like you can plugins are only

1:17:06available in CoWork. Okay, so I'm in

1:17:08chat. Let me go back to CoWork.

1:17:10And

1:17:11uh

1:17:12customize. Here we go. So, they they

1:17:15moved the skills into plugins. So, for

1:17:18example, I just installed this legal

1:17:19plugin or you can hit plus and you can

1:17:22browse plugins. This is what for my last

1:17:24video that triggered the SAS apocalypse

1:17:26is

1:17:27they they they they created a whole

1:17:29bunch of legal skills and just made it

1:17:31free to the world. Anybody could use

1:17:33them and we saw 20% drop in the stock

1:17:35market for legal

1:17:36publicly traded companies in 1 day. And

1:17:39they've been rolling out more and more

1:17:40of these skill packs. So, within CoWork,

1:17:42this is where you can see them.

1:17:44And if you sort of browse through the

1:17:46plugins, you can say, "Oh, I want this

1:17:48marketing or legal or whatever." Right?

1:17:51And so, getting the skills is is really

1:17:53easy. Um just be careful if you are

1:17:57working off of these open source

1:17:58platforms like Open Claw or Hermes and

1:18:02you're getting your skills off the web.

1:18:03A common security attack point is that

1:18:06they'll create fake skills and people

1:18:08download them and then it's actually

1:18:09malicious um within the skill. But yeah,

1:18:12like if I install this legal pack, which

1:18:14I did, um that just means that when I go

1:18:17to CoWork and I use it, I can then ask

1:18:21it to bring up my legal skills. Um

1:18:27Oh, they're now under plugins. Plugins,

1:18:28legal. So, if I'm going to do a

1:18:32review a contract, I will click that and

1:18:35that will boot up the review contract

1:18:37skill. And what is a skill? It's really

1:18:39just a prompt. Like if you look behind

1:18:41the scenes of any of my skills that I've

1:18:44built,

1:18:45um

1:18:46Like here's my requirements agent skill.

1:18:50It's just a prompt

1:18:51telling me what to do or telling the

1:18:53agent what to do. And so, when I call

1:18:56the requirements skill,

1:18:58which I usually do through my platform,

1:18:59I'll just call a requirements agent. But

1:19:01in Claude CoWork world, if I wanted to

1:19:05do my requirement requirements gathering

1:19:07with a client here, I would just grab my

1:19:09requirements agent skill

1:19:11and then

1:19:14help me

1:19:15run a requirements

1:19:17session with a client and it will

1:19:19instantly know what to do because it's

1:19:21booting up all of with the skill. So,

1:19:23Co-work is a great place to start. They

1:19:24have a massive skill library. It's

1:19:25recently moved over to plugins.

1:19:28Um if you're researching, you can build

1:19:30your own skills very, very quickly. And

1:19:32just like prompting in the past, like

1:19:34when you build a when you build

1:19:35something from a prompt, you want to

1:19:37look at the output of your prompt and

1:19:38say was it good? If it wasn't good, go

1:19:40back and then say what was wrong with it

1:19:42and do it again and do it again. And so,

1:19:44prompt engineering still applies to

1:19:46this. Uh so, when you create skills,

1:19:48you're really just engineering prompts.

1:19:50You're storing them, you're saving them,

1:19:51and you have the ability to call them by

1:19:53name. So, when I say build a custom

1:19:54sales proposal, when I say those words,

1:19:57it will automatically invoke my sales

1:19:58proposal skill within Claude, which I've

1:20:01already defined.

1:20:02Um so, yeah, Claude's a great way to get

1:20:04started. Skills are essentially open

1:20:06source. We give them away to our

1:20:07clients. We're constantly improving

1:20:09them. I don't think the value is so much

1:20:10in the skill. I think the value is in

1:20:12the person behind the skill and how

1:20:14quickly they're iterating on it. So, if

1:20:16you go to managed agents um and you have

1:20:19human in the loop expert say behind your

1:20:21skill for SEO, like we have an SEO

1:20:23manager, they're constantly improving

1:20:25that. So, if you're doing them on your

1:20:27own, you have to be like I I when I was

1:20:30building them on my own, I had a skill

1:20:31that would literally do regular research

1:20:34for best practices for that specific

1:20:36skill, would compare the research with

1:20:38what I was doing, and then make

1:20:39suggestions on how I could improve my

1:20:41skills. But skills should not be static.

1:20:43They they should they need to be uh

1:20:44evolving. And there's a ton of open

1:20:46source out there. So, throwing a lot at

1:20:48you there. Hopefully that answered some

1:20:49of your questions. Probably created a

1:20:50few new ones along the way as well.

1:20:56Yeah.

1:20:57Don't you wish they would just all nail

1:20:59their terminology? It's kind of like

1:21:015 years ago, couldn't just Samsung and

1:21:02Apple get their freaking adapters the

1:21:05same here? Like

1:21:07GP custom GPTs over here, it's a gem and

1:21:09Gemini, and it's a skill over like

1:21:12Call them the same things.

1:21:17All right, we got 7 minutes left for

1:21:18questions.

1:21:23Like

1:21:25What's the difference between doing a

1:21:26project through the chat with the V, the

1:21:29the the co-work?

1:21:33Doing a project using Claude co-work or

1:21:34doing a project through your voice broke

1:21:37up a little bit there.

1:21:38Oh, so chat is um not an agent. Uh chat

1:21:42is just chat. So,

1:21:44Everyone's been flooding over to Claude

1:21:46right now. Their revenue has just

1:21:48surpassed chat GPT. It looks like their

1:21:50next funding round will be even a higher

1:21:52valuation than OpenAI. Uh everyone's

1:21:55been gravitating over to Claude for

1:21:59their agents, Claude co-work and Claude

1:22:01code. And Claude co-work being Claude

1:22:03code for non-technical people or I say

1:22:05Claude code for dummies. Claude co-work

1:22:07is very, very powerful. But if you just

1:22:09go to Claude chat, it's no different

1:22:11than chat GPT or Gemini from a chat

1:22:13perspective. That's not the the the big

1:22:16breakthrough. It's Claude co-work

1:22:18running on your desktop is the first

1:22:21real usable high volume agent other than

1:22:23you maybe last year. Mana's came out

1:22:24very expensive, but um Claude co-work,

1:22:27the desktop agent, that's what we're

1:22:29talking about. Claude chat is just

1:22:32another chatbot like chat GPT.

1:22:36So, but yeah, this this talk we're we

1:22:38were going a bit past that. May like we

1:22:41might want to do a workshop to go just

1:22:43deeper into Claude co-work for those of

1:22:46you that are in that space. But when you

1:22:47graduate from Claude co-work, you'll

1:22:49move into multi-agent orchestration. So,

1:22:52like understand the stages and don't try

1:22:54to get ahead. Like if you haven't even

1:22:55learned the basics of prompting and how

1:22:58to use chat GPT back in the day and

1:22:59you're getting straight into co-work um

1:23:02or orchestration, like you need to go

1:23:04back and learn how to prompt and learn

1:23:05the basics and, you know, work your way

1:23:07up that that ladder. From my my previous

1:23:10talk, I had a

1:23:13a visual representation of of that

1:23:15ladder. I think it's important that you

1:23:16don't get too far ahead of where your

1:23:18current skills are and really overwhelm

1:23:20yourself.

1:23:23But yeah, don't play with Claude chat. I

1:23:25mean, Claude chat is fine, but it's not

1:23:27an agent.

1:23:34This is the sort of maturity ladder I

1:23:36would call it of This is from the

1:23:38workshop last month. This is the part

1:23:41two of

1:23:42is this

1:23:47Right? So,

1:23:49learning how to chat with agents out

1:23:51here, chat GPT, just using it, becoming

1:23:55a power user, learning how to connect,

1:23:58build custom GPTs or, you know, using

1:24:01projects and connectors and

1:24:03custom instructions and all of that.

1:24:06>> [clears throat]

1:24:06>> Most people think they're good at chat

1:24:08GPT or AI chatbots, and I sit down with

1:24:12them and they usually suck.

1:24:14There's a lot of depth to this. So,

1:24:16becoming a power user of the basic chat

1:24:19tools

1:24:20is the first, you know, couple steps.

1:24:22The next one is mastering your first set

1:24:24of agents. Co-work agent is a great

1:24:26place to start or if you're technical

1:24:28and you do it right, playing around with

1:24:31um

1:24:33Codex also on this side.

1:24:35And then progressing into the

1:24:37multi-agent teams from here. That's the

1:24:39either open source, open claw, Hermes,

1:24:41or platform like AI 1 or Perplexity

1:24:43computer. Getting over here.

1:24:46But yeah, in each of these steps, like

1:24:48they're they're big jumps.

1:24:50Mastering co-work, I I've put in

1:24:52hundreds of hours and I keep finding

1:24:54more and more things that I can do here.

1:24:55And it's constantly changing. Like they

1:24:56just roll out co-work design and now I

1:24:58need another 10 hours just to master

1:25:00that piece.

1:25:01Yep.

1:25:04There is a a question from Amanda.

1:25:06Amanda.

1:25:07Hi. Thank you first for this because I

1:25:10find this very helpful, especially since

1:25:12it gives me context and a visual view of

1:25:14how it actually looks and be able to

1:25:15build it. Uh I did have a question in

1:25:17regards to the different models. So, I I

1:25:19understand that with no co-work is where

1:25:21everybody is moving on right now. I've

1:25:23been using it myself. Um but I also have

1:25:26found that because certain organizations

1:25:28and where we're working, um other people

1:25:30use uh the Microsoft version. But have

1:25:33you had to be able to port your agents

1:25:36over from one platform to the next

1:25:38because as we know, things change very

1:25:40quickly at a very high rate. Um and I

1:25:43was curious and know we spend all this

1:25:44time building this. Um if you've had to

1:25:47port any of your agents over to

1:25:48different platforms.

1:25:50Beautiful question. Thank you for asking

1:25:52that. It it is one of the top questions

1:25:54that we hear when people try to migrate

1:25:57from one platform to the next. I I made

1:25:59a post on it in in that community just a

1:26:01few days ago. It's it's so important.

1:26:03So, first of all, Microsoft is not even

1:26:06in the top five right now. And you have

1:26:08to be careful what you build on

1:26:09Microsoft. Just look at Google share

1:26:10price versus Microsoft over the last 6

1:26:12months. I think the spread is about 45%

1:26:15difference. Like Google is doing this,

1:26:16Microsoft is doing that. Markets are

1:26:19forward-looking. So, the the the stock

1:26:21markets are saying Microsoft isn't super

1:26:23well positioned for AI right now. They

1:26:25might turn it around. I wouldn't rule

1:26:26them out. They're still worth $4

1:26:29trillion and they can probably turn it

1:26:31around. But when you build on top of a

1:26:33platform like Microsoft, you are locking

1:26:37yourselves into their their

1:26:38infrastructure just like you would say

1:26:40if you started building a bunch of

1:26:41customizations on top of salesforce.com.

1:26:43So, the way we recommend that people do

1:26:45that first of all is whatever you build

1:26:47should be platform agnostic. Like you

1:26:50you want to connect to say Microsoft

1:26:52with a multi-agent orchestration and

1:26:55just plug in with API. Like just a

1:26:56plugin from their data systems to your

1:26:59agents. So, you can build agents around

1:27:01Microsoft. And so, the reason why this

1:27:04this this is so important cuz if you

1:27:06want to suddenly switch all of your

1:27:07intelligence from Microsoft or all all

1:27:10you want to switch from Microsoft to say

1:27:11Gemini or Claude or Grok 5, maybe Elon

1:27:15comes out with something and you hit a

1:27:16drop down and your whole platform

1:27:19switches to the next LLM. So, that

1:27:22design is really important. And the way

1:27:24that you port your your content over

1:27:28from one platform to the other is first

1:27:29of all, don't have it nested and

1:27:31embedded with within Microsoft. When

1:27:33it's external to it, then you have all

1:27:35of your skills, your data, your brain,

1:27:37your recipes, your workflows, your jobs,

1:27:38like all of those assets, you can take

1:27:41them with you. What I I recommend my

1:27:43clients do um is they can they can leave

1:27:46me at any time. Like it's just a bunch

1:27:47of GitHub files, right? So, like if if

1:27:50It's funny cuz we're doing software

1:27:51development here and we don't even

1:27:52realize we're doing it cuz we've

1:27:53abstracted away the layer of software

1:27:55development. But those of you that know

1:27:56what the GitHub is, like if I look at

1:27:59this is the the code repository.

1:28:02Um if I look at all of my agents from a

1:28:04GitHub perspective, these are my skills.

1:28:07And my skills are just markdown files.

1:28:09Like here's my skill for optimizing for

1:28:12AEO.

1:28:13This is it. This is all the code. And I

1:28:16can take it with me. I could I I'd say

1:28:19organize your If you want to do it

1:28:21proper, keep all of your agents external

1:28:24to the major platforms. Like I wouldn't

1:28:26even build on top of Claude agent teams.

1:28:29It's it's a really powerful tool, but

1:28:30you're locked into Claude. And we just

1:28:32saw a GPT 5.5 come out and it could be

1:28:35better. And suddenly like, ah, how do I

1:28:36get all my Claude agents over back over

1:28:38to chat GPT and now over to Gemini? Have

1:28:41them in the middle.

1:28:42You should be able to, you know, switch

1:28:45models with one drop down. Have all of

1:28:48your assets, your skills, your agents,

1:28:53your data, all backed up on your own

1:28:56server and you can take it. Like you

1:28:58could literally go from Myzone AI 1 over

1:29:00to Perplexity computer or to open claw

1:29:01or Hermes, grab all those files and say

1:29:05process them. And so so if but you have

1:29:08to remember that you need to control and

1:29:10own all that data. I don't know if all

1:29:12the major platforms will allow that

1:29:15portability, but I would demand it from

1:29:16whatever platform you're going to go on.

1:29:18Okay, yeah, that I think that that helps

1:29:20a lot cuz I yeah, I think the biggest

1:29:22concern with a lot of the companies

1:29:23especially that I'm working with is that

1:29:25they don't necessarily trust one one

1:29:27platform over another and mostly because

1:29:29of security is why they have chosen

1:29:31Microsoft. It's not to say it is the

1:29:32best, but it's more out of uh

1:29:35education obviously for one because I

1:29:37don't think any of us truly know what

1:29:39capabilities are are there yet um

1:29:42unless you're in it like you are. And so

1:29:44a lot of people are just unfamiliar with

1:29:46which ones and they're just going with

1:29:47whatever their IT teams are willing to

1:29:50let them use.

1:29:52Yeah, it's it's crazy. I I see a lot of

1:29:55big companies are like, "Oh, we don't

1:29:57have permission to to use this tool or

1:29:59use that." And I'm like, "Oh my

1:30:01goodness, you guys are going into a a

1:30:03nuclear war with

1:30:05like

1:30:06sticks.

1:30:07Some of them they're just locked down

1:30:09from

1:30:10protocol and process and whereas

1:30:12entrepreneurs are flying. Like you you

1:30:14need the ability to just hop on the

1:30:16right model and and go in. These windows

1:30:18are huge opportunities, but they're

1:30:20closing fast. So I think a lot of these

1:30:22big companies that have those sorts of

1:30:25uh restrictions and limitations and,

1:30:27"Oh, we'll review that in our next board

1:30:28meeting." I I don't think they're going

1:30:29to survive to be honest. Like the the

1:30:30world everything is changing right now.

1:30:33So if you're if you're locked into

1:30:34Microsoft right now, it could be it

1:30:36could be like the the death nail in your

1:30:38business model over the next 2 years if

1:30:41Gemini absolutely blows up and crushes

1:30:43Microsoft, which is entirely possible

1:30:45right now. Based on the Google share

1:30:46price,

1:30:47that's what the market is predicting.

1:30:49Yeah, that's fair. Yeah.

1:30:52Cool. Well, I I had a 10:30 to 11:00

1:30:54meeting. It just canceled, which I'm

1:30:56very grateful for. So if anybody wants

1:30:57to hang out for another little bit

1:30:58longer, you guys know me. I just love

1:31:00talking AI. I learn a lot from the

1:31:02questions as well. Um I'm good to keep

1:31:04asking answering questions and debating

1:31:07if anybody wants to.

1:31:14All right, looks like we're wrapping up.

1:31:15So good to see you all. Hope to see you

1:31:17next month for our two workshops, one on

1:31:19security, one on token optimization.

1:31:22Dive in, have fun, pace yourself. It's a

1:31:24marathon, it's not a sprint.

1:31:26Be careful with security if you're going

1:31:28to go off and do this stuff on your own,

1:31:30but have fun. You learn the most just by

1:31:32getting your hands dirty, diving in

1:31:33there. And I really think that this is

1:31:35important for every entrepreneur to know

1:31:38in 2026. So have a great rest of your

1:31:40week, guys. We'll see you on the next

1:31:42workshop.

1:31:43Oh, Brian.

1:31:44>> Hey, Mike.

1:31:44>> Yeah. Yeah. Yeah, I think Janette I

1:31:46think Janette's got it. Go ahead,

1:31:47Janette.

1:31:48Yeah, there was a question here. What

1:31:49are your top two to three YouTube

1:31:51channels on AI?

1:31:53My zone AI is incredible. Yeah.

1:31:57>> [laughter]

1:31:58>> Oh, wait, we got to see the new website.

1:32:00Uh it just finished. All right, let's

1:32:02check out the new website for

1:32:05uh there's the new Lux Decor website.

1:32:08So we turn houses into homes. Well,

1:32:09first of all,

1:32:11so now what is your actual website

1:32:12address right now?

1:32:14Uh W My actual address is lux-decor.com.

1:32:18Yeah. Okay, okay. So

1:32:20um

1:32:21And now first version, we don't have any

1:32:22imagery yet. So that's where I usually

1:32:24bring in Nano Banana. So I would still

1:32:26have one more

1:32:28uh it already I guess it stole some

1:32:29images from your site.

1:32:31No, those are not my images. I don't

1:32:32know where those images came from.

1:32:33>> Oh, it it probably created them with

1:32:35Nano Banana. That's crazy though. So

1:32:38this is your current site. Portfolio,

1:32:39services, about us, blog, a home that

1:32:42reflects you. I remember the S this CRO

1:32:45report was talking something about the

1:32:47benefits. Like it didn't like your

1:32:49uh benefit statement and trust signals

1:32:51above the fold. Let's see how it

1:32:54um

1:32:55Wait, where did

1:32:56Oh, here it is.

1:32:58Let's close some browsers.

1:33:03We turn houses into homes. A studio for

1:33:06today. And look at the calls to actions,

1:33:08right? We've got this is a CRO thing.

1:33:10Like do you do you have clear call to

1:33:12action above the fold? You got to get

1:33:13started.

1:33:15But

1:33:15>> That's That's a discovery call and on

1:33:17the top, too. But like book a discovery

1:33:19call or see the work. Like it's And do

1:33:21we have any trust signals above the

1:33:23fold? You don't. Critical for CRO and

1:33:25you can see here how it's saying like

1:33:274.9, 312 client reviews.

1:33:30Best of multiple years. That alone can

1:33:33do huge. Also trust signal established

1:33:352003.

1:33:37Um yeah, so you can see it made an

1:33:39effort for trust signals and clear calls

1:33:42to action. Uh it even did the French and

1:33:44English. I wonder if it built out to

1:33:47build out the whole site. Oh, it did.

1:33:49Here's your portfolio pages,

1:33:51your shop.

1:33:53Oh, looks like it's still working on

1:33:55some things. Sure it'll done.

1:33:59That's incredible. Uh it looks like it

1:34:00just created the first couple of pages.

1:34:01So services pages we have

1:34:03No, it was the home and

1:34:10And also changed the services

1:34:12completely. That's hilarious.

1:34:14Wow.

1:34:14>> Oh, really? Yeah.

1:34:16Yeah, and and that's a lot of it was

1:34:18from all that research. What what are

1:34:20your competitors doing? And analyzing it

1:34:22from different uh perspectives.