Build agents with skills and automate recurring tasks

For an infrequent over the summer episode of the AI Academy. Steve and Dustin here welcoming you one and all. Dustin and I were just catching up. Dustin was just finishing up his summer holidays. He was at the beach and in Las Vegas with the family. I have been struggling trying to get to our property in the interior of British Columbia, but the forest fires have other plans for us. Mm.

So the air quality at the lake that my family's at, Dustin, the other day the recommendation was healthy people can stay outside for 20 minutes. If you have had any issues with your heart or lungs or anything, do not go outside. Yikes. Do not go outside. It was 11. I don't know what the scale, I don't know if the air quality indexes have different scales, but it was 11 or 12, which is the highest that I've ever seen.

Well, I guess, no, it has been worse, but it was still high and holding at high. So at any rate, I hope everybody is safe and happy and not affected too much by the ravages of mother nature this summer. And we welcome you back. Dustin. Good to be back. I do miss the beach because it was nice and cool there.

It is very hot here, and we're struggling through this August heat, but temperature's going to go down, so it'll be- And is this the hottest month for you? It's usually July and August are the hottest. Yeah. And you just got to endure, get to September, and hopefully it'll be mid 80s and more endurable. There you go. All right. Well, so what are we going to talk about today?

Yeah, so lots happened. Over the summer, we've been shipping like crazy. Our team has just been working super hard to transition the app to a more hands-off AI assistant, and the way that we have rolled this out, several weeks ago at the Academy, we talked about scheduled tasks, and everyone who was there probably remembers we went through how to schedule a task one time and how to create recurring tasks that happen over and over again automatically, autonomous tasks.

But between then and now, what we've done is we've given a major overhaul to personas. And for those of you that have been around and have heard me talk about this, you know that personas were something that Magai had from day one back in March of 2023, a whole year at least before OpenAI invented what they call GPTs. And Steve, it still drives me crazy today. Just as early as two weeks ago, I think, somebody asked me, "Does Magai have GPTs?"

And I had to say, "Friend, we had them before OpenAI did. We just called them personas." We had them a year before they did. So we've now transitioned personas, we've essentially labeled them differently because the industry is changing, the terminology is being solidified in people's minds, and now the terminology that is being used for a specialized version of the AI model that does a specific task is now agent.

And this has been adopted throughout the industry, so we thought it would be a good time to sort of wave goodbye to the terminology of persona because some people still struggle with the idea. But many people recognize the idea and concept of an AI agent. So personas are now called agents inside of Magai. But it wasn't just a label change. We're not just changing the name of something.

We actually added a fair bit of capability to these personas to enhance what they can do and to keep up with, again, the growing industry standards that are happening. If you are in the room and you've heard of the concept of Claude skills, throw a little hand emoji in the chat. We've essentially now given these agents the ability to have skills. Now, what is a skill?

Essentially, a skill is nothing more than an instruction, but allowing agents to have more than one instruction, whereas previously... Let me just start sharing my screen here. We'll dive right into it. Previously in Magai, you click on a persona and you-- Let me see if I can find one, like Bruce. You hit the settings, and you have the persona's or agent's name right here. You can give them a name.

You can give them a profile picture. You can give a little description that just describes what they're doing. And then you give them instructions, right? And they follow these instructions. This defines their personality. This defines how they interact with you, right? And it defines the tasks and the specialties that you want to be sort of front and center when you interact with that agent.

And that was sort of the only way to build custom agents. Well, skills brought in this idea that maybe these agents need to have multiple different things that they specialize in for different pieces of the task, allowing you to be more elaborate with what those instructions entail. Maybe it's entire operating procedures. Maybe it is a source of information or knowledge or a source of truth. Maybe you want to have several different skills built in rather than having to have an agent for every different skill,

right? You can sort of bundle things together. For example, you can have a marketer agent that has the social media marketing skill and the SEO skill and the press release writing skill. You can have multiple different skills rather than having to break that into multiple different agents. Does that make sense, Steve?

I think so, yeah. Okay. So we've added the ability for you to add skills to these agents. And so you'll see a section right beneath the instructions here where it'll display any skills that you've added, and you can add skills. You can write them manually if you are comfortable writing skills. Simply a name for the skill, a description, and then those instructions, just like you would give to a persona previously. Or you can actually drop skill files in here.

So throughout the internet, people, since Claude sort of standardized this idea of skills, people have been sharing their Claude skills on the internet and their agent skills everywhere. You can find these big marketplaces and repositories where skills are freely downloadable. So what you can do is you can find a Claude skill that works really well for someone else. You can drop it right into Magai, and now your Magai agent has those Claude skills.

Some examples I can give here is I created a PowerPoint creator. And what I did is I found a skill that somebody had shared that was a PowerPoint generation skill, and this is a more complex one. It's not just text instructions, but this one actually has multiple files. It has design guidelines, it has a preview template, it has a schema file.

It has some code files to help it execute some code. So these skills can be much more elaborate and much more sophisticated in helping you do more complex tasks. And so having that ability in Magai to adopt the industry standard that is now becoming more widely practiced, you now have the ability to apply any of these skills being used in any of these other agents, any of these other platforms, and drop them right into Magai and create your own custom agents. Now to make that even better-

This reminds me of the... It's kind of like the helicopter pilot in "Matrix." "I've got to learn to fly this helicopter." Yeah. Basically, you're just uploading the skill set. "I know kung fu." Yeah, now your agent knows kung fu. It's crazy. So on top of that, Magai now has the ability to create agents and skills for you.

No longer do you have to switch over to the persona generator to generate a persona and then copy the instructions, and then open the persona window, create a new... None of that. You just say, "I need a sales copy writing agent," a new one, because I probably already have one in here. "Specializing in," oh, let's just say CrossFit, "CrossFit marketing.

Create the agent and create the skills necessary for this agent to create all kinds of sales copy for me." I can't spell. There are many typos, but the AI will get it because it doesn't care about our typos.

So watch this. I'm going to send this off. I'm just using the assistant agent by default, and it's going to do all of this work for me. Okay, so it's going to think about it, and so it has now seen that it has the ability. Claude is looking through the tools that we give it in Magai, and it sees that it has the ability to create its own agents.

So it's going to first create the agent inside of your workspace. And there we go. CrossFit sales copywriter agent has been created. Now it's going to build out its skill set. I probably could have picked a much faster model to run this on for demonstration purposes, but ...

Well, and you also have, as you expressed to me earlier, your internet is basically dial-up modem speeds. Yeah. One of the things I was telling Steve before this is our trip to the beach was great because it had internet speeds that were four times faster than my home internet, which is usually not what you get at an Airbnb. So it was a pleasant surprise.

It's amazing how quickly you get velocitized to the new speed, isn't it? Right. Yeah. Everything's just so fast. Like two days later, you're not even noticing it until it's slow again. Great. Okay. So we've created the first skill here, CrossFit offer and avatar brief, and we can even view that skill if we want. It'll open up the agent editing panel, so we can see the instructions that it created for this agent.

All of this we didn't have to write, which is great. And then here's our CrossFit offer and avatar brief. Use first on any new CrossFit copywriting project or when the offer, audience, price, or deadline is unclear. Builds the intake brief, member avatar, offer structure, and message angles that every other copy asset is built from. So let me ask a question so I can start to put everything in its place to create a little more of a 3D model in my mind and hopefully in the rest of our minds.

Sure. So these are the skills and the kind of the processes that the agent goes through. Yeah. The agent can be then inserted into any workspace which will know the location of your business, what your membership rates are, who your target audience is. Mm-hmm. That will live in your workspace still. You don't ingest those skills into the agent, but instead the agent has more generic skills that it'll pull from your workspace in order to create when you're actually going, starting to create your specific content.

Have I got the Hierarchy correct? Yeah. Let me see if I could just reframe it a bit to make sure that I understood it correctly. So as every Magai user should know, your workspace dictates a lot of the context that your AI understands about you. So your workspace should contain specific details, custom context about your business, your products, services- Your pricing- Pricing ... the products that you sell- All of it ... your location.

Right. Yes. Your value proposition. Just like in Claude or in ChatGPT, you have projects. We have workspaces. And this workspace is where the majority of your baseline context for your business or your whatever project you're working on is. And so your agent is already going to have access to that information. These skills are added to the agent to give it more specific workflows, more specific expertise, and refinement of whatever objective you have given to that skill. And it's going to be augmented with whatever context your workspace has.

And here's the really cool thing about skills. So thinking of usage consumption, right? Every bit of text that the AI has to process consumes usage, and I think we talked a couple weeks ago about token consumption and how that all works. What makes skills really great is that the agent isn't loading every skill every time.

And so if we look at this one in particular, this particular skill has a long list or, well, a short list of instructions here, but then it's also got files attached which contain more instructions. And there's going to be some improvements to this view, so you can actually view these markdown files, more specifically, make edits to them. This is sort of version one of this. But all of this is text that's now added to the context of the conversation, consuming more tokens, right?

So more elaborate skills means more usage consumed. However, when you have broken this up into multiple skills, it's not like the agent is loading every single file of every single skill. What it's able to do is only load what it needs to load. So if the goal is simply just paid social ads, it's only going to load that skill.

It's not going to load unnecessary context, unnecessary expertise or anything. It's only going to load that paid social ads skill, saving you on usage and making it the most efficient use of your output as possible. So we can see here that Claude really went above and beyond. It really wanted to make this sales copywriter elaborate. How many skills did it create here?

Like seven skills? All for individual different purposes. So we've got an organic social content skill, retention and referral copy, SMS DM sales scripts, email sequences, paid social ads. We have a very detailed agent now for our CrossFit sales copywriting. And again, we can go in and we can review all these.

Now you've got there Switch to Agent button. So do you actually have to select which one you want to use if you're creating a project? Yeah. You're creating content. Good call-out here. So the Switch to Agent button is simply if you want to just switch to that agent in the chat. So notice how it changed here down in the prompt box. So if you were creating content, if you were starting to create a marketing campaign, and you wanted to create a global campaign, you wouldn't select that. You would let it select which agent it is.

But if you were actually going in and say, "I want to create a series of Facebook posts," you would then choose your social media agent. Is that the process? Well, so the skills are attached to the agent. So skills aren't individual things that you activate. You activate agents.

Okay. So I pick my CrossFit sales copywriter, and that copywriter is equipped to do any of these things. Okay. Any of these skills it can have and call upon when the task requires it. Okay. So again, if I open this up... Oh, sorry. If I view skill. So this is the agent we switch to when we want sales copy for our CrossFit business.

Okay. We choose this agent, and no matter what that task is, this agent now has a high-level skill to complete that task. So if I want to do, for example, organic social content, all I need to do is make sure that agent is selected and say, "Now give some organic social content ideas for our new CrossFit gym."

And so what you'll see here is the agent is going to load that skill. It's going to say, "Hey, okay. Among my list of skills, which one of these do I need to perform this task?" Thought for a second, loaded agent skill, CrossFit organic social content. It's going to read some files. It's going to make sure that he's got all those skills files loaded so that he can do that job, or it can do that job as effectively as possible.

Does that make sense? I think so. Yeah. So think of it like, again, instead of having to create a different agent for literally every teeny tiny task, you're creating one agent that has multiple abilities, that has multiple skills- Mm-hmm ... that are refined per task.

Now, when Claude talks about its skills, one of the things that it talks about is the strengths is it says it can do multiple steps, so that it can do sequences. Mm-hmm. How does that fold into this? Yeah, it's the same. So every AI agent can do multiple steps these days. They're really, really good, and I'll show that with one of our examples here today. I'm going to switch over to our one that we did, or one that I did late Monday night, I believe it was. Monday or Sunday night.

And tell me if you've done this before. You're laying in bed late at night and you're like, "Oh," just scrolling through threads or Facebook or Twitter, and I see something that somebody did with AI and I'm like, "Oh, shoot. I want to go try that." And then you just go try it right then and there. Anybody else do that? I do it frequently. And for me, it tends to be X. That just happens to be the place where, I don't know what happened with the X algorithm, and everybody says X is terrible. For me, it gives me literally only the best AI stuff.

And I have to not go on it now because I'm like, every time I go on there, I go down an AI rabbit trail of something cool that I want to try out. That's just how the algorithm has worked for me. So I found this tweet from a friend of mine named Tim, who owns a social scheduling tool company. And I'm going to point out or reiterate, I was in bed half asleep. I found this tweet and I was like, "I want to try this in Magai."

And you can tell that I was tired because instead of saying, "I found a tweet," I typed, "I foundation tweet" from someone that said this. Using your search console connection, do what he's saying for Magai. So here's what he said in his tweet. I literally just copied and pasted. He said, "Easy SEO win. Go to Google Search Console, Performance, set the date to 12 months, then export it.

Upload pages.csv," which is the file you get from the export, "and queries.csv," another file you get from the export, "to any AI tool. Tell it to find queries you rank for but don't have content for. Go write the content and you will rank quick." And so knowing what I know about Magai's capabilities, I'm like, "Oh, this is way easier in Magai." You don't have to do all these steps.

So he's saying, go to the search console yourself, you do the thing, you set the dates, you export, you grab those files, and then you go take those files and upload them to an AI tool, and then tell it this, and then the tool will tell you what to write in order to rank better. And then you can go write that content yourself. I'm like, well, Magai can do pretty much all of this without you doing anything.

So I gave it that tweet. Magai now, as of before the summer started, we have integrations. And one of these integrations is Google Search Console. So if you have a Google Search Console account, you can connect your Google Search Console to Magai, and any AI agent you use can then access your Google Search Console directly without you having to go and manually export and then import files. You can just tell it, "Go do this," and it'll go find the data.

So that's what I did. I pasted that tweet and then I gave it Magai's site map, not Magai's silent. Again, I was tired. So Magai's site map is at this URL. I pasted the URL. So here's what Claude Opus did. Pulled search console data because it already had the connection. I do have one question for you. How do the agents choose the model to use?

The agent uses whatever model you currently have selected. Okay. Yeah. And so now the other thing I did here is I did have my marketing agent already activated. Yeah. So it did have some specialization there. So you had an agent similar to the one that you just showed us being constructed. It was who you- Actually, no ... would have been instructed through.

Let me look at our Magai marketing strategist. I don't think it has any skills. Yeah, this was like a persona that I- Okay ... created a while ago. So not even a very sophisticated agent, because I didn't really need the agent to be super specialized because number one, this was an experiment. This was more about research than it was about content. Yeah. This was more research and data analysis.

So anyways, here's what it did. It fetched my site map. Well, it failed the first time, and then it looked up the Google Analytics, the search query data. Pulling 12 months of query page data now plus query page pairs so I can see the actual gap. So it did another tool, pulled all this data, and then it pulled 30,000 queries and 1,136 pages. And now, then it mapped each query to the page Google actually ranks, and grabbing my site map again.

Seems like it failed again for some reason. And then it split the task into smaller batches because it was a lot of data. Used another tool. And you can see it's kind of working through this problem step-by-step systematically. And we can open these up and see what it's thinking about. We can open up its thoughts here. And we can just see all the tools that it's using. And here's what it finally did. It said, "Done.

The full audit is in the canvas." So it built me a canvas document. Here's the Magai content gap audit, 12 months of search console data. And it kind of put this into a very elaborate and very comprehensive report for me. It built priorities of what we should do first and what we should do second, third, fourth, fifth.

It gave me very detailed instructions here. I didn't even read through all this because I'm like, that's a lot of reading and it's late at night and I don't want to read all of it. So, I kind of just scanned through all of it And then it gave me some suggestions here, and then it gave me the one metric to measure. And how to measure if everything that it built here, if we implement it, how do we measure the success?

Then it gave me some suggestions here, which is a new feature you might notice in Magai. It'll actually give you next step suggestions, little buttons that you can click to keep the conversation going. So I didn't want to stop there because in my head I'm thinking, "Well, this is great, but I don't want to have to go read all this." Our team uses Todoist as our project management and task management tool.

And then I have my two VAs on there who that's where they get their tasks from, and we monitor task and projects. So I told Claude, because I knew that it also had a connected integration to Todoist. My Todoist is already connected here. I said, "Set up the Todoist structure now..." Or sorry, "Create a new SEO project in Todoist and a task for each item on the list you created. Keep the project well-organized so we can knock things out efficiently and delegate what can be delegated to teammates."

I didn't tell it what teammates. This fairly dumb prompt, right? Not super elaborate. Again, it's late, I'm lazy, I'm tired. So I just wanted it done. So Claude did that. It said, "I'll set up the Todoist structure now." It found the Todoist tool and then what I was connected as. It loaded all of the projects in Todoist to see if there was any current SEO project. It found an existing SEO project and decided, well, I'm going to create a new one.

So it created a new SEO project. Project was created and it also found my team roster, the team members, my VA Precious and my other VA, Vilni. And it started assigning tasks to everyone. You can see it really did a lot of work here. This thing must have worked for five minutes to build all this.

And it told me here the structure and all this information, but I'm like, "That's great. I'm not going to read all this." I just opened up Todoist and are you able to see my full screen here? Can you see Todoist? I don't see it right now. We're still on the AI screen. Let me change what I'm sharing here. You have to share- I'm going to share my screen I had issues sharing full screen the other day, but maybe it's just my computer.

Okay. Can you see Todoist now? We do have Todoist. All right. So here it is. It created this project, the SEO Content Gap Sprint. Gave it a title, gave it a nice description, which includes what our target is, and it created every one of these individual tasks. It assigned specific tasks to specific people.

And here's what really blew me away. It created a section called Start Here, and it has an onboarding task for each of my VAs so that they know exactly what to do and how the project is. So it has basically given them the context for the project. Right. So they know exactly what the project is, their role in the project, what the objectives are. And so when they come in here and look at all these tasks, they have sort of their direction.

I don't have to give them that direction. Assigned me some tasks, too, and then it broke everything out into its own section, prioritized each one, gave them due dates. Right? You can see all these... It was just so elaborately done that I was blown away. And after that, let me go back over here to Magai. After I sort of reviewed that Todoist task...

Let's see, where was it? Oh, there it is. I reviewed it and I noticed it gave one of my VAs a task around building web pages, and that's not the type of thing I give her. So I kind of just corrected. I said, "Precious is not a builder. Both Precious and Vilni should only have tasks that don't require technical knowledge. They can use Magai to help them split up the work, and actually, they'll probably need agents with appropriate skills here in Magai to help them accomplish these tasks at the highest possible level. If we don't already have agents with skills appropriate for these

tasks, create them for me. Assign all the technical building stuff to me." So you are giving much broader instructions than we originally, like six months ago, we would be comfortable with. Right. Because six months ago, Magai couldn't create its own personas. Yeah. You would've- It couldn't connect to third-party tools.

Yeah. And so all of that technology has been built to now that agents are... Well, now that AI models are capable of connecting to multiple tools and having much more elaborate nuance to the tasks they can do. And now that they can work agentically and not just do one output with maximum of... I remember the first time Claude produced content that was like 8,000 words at once.

That was mind-blowing. It wasn't possible last year. Yeah. At most, these AI models would output 3,000 words, 4,000 words, 5,000 words if you really pushed it. But the first time it generated 8,000 words of content all at once was mind-blowing. But now, that's nothing, right? Because they can work multiple tasks, multiple sequences. They can work for minutes at a time.

Some really powerful models can work for hours at a time, just working and chipping away at a task. And so these models have reached the point of capability that we can now hand them very elaborate tasks. Good. Such as building entire projects, doing the research, doing the analysis. Yeah, we would've broken it down into... We would've had to break it down into chunks or it would've gotten completely off the rails.

Right. We would have had to step-by-step ourselves through it. Yeah. So here's how it sort of did this next bit of it. So I wanted it to create the agents that would help my team accomplish these tasks. So it went and it searched through our owned agents, what agents already existed, and then found no existing agent has SEO skills attached. So it pulled the project again, figured out what agent and what skills that agent needed. Project confirmed. Okay, great.

It created the SEO landing page copywriter. It created the SEO CTR rescue editor. It created the AI comparison content writer. All of these are agents for specific tasks throughout this very large 48-task project. There are a total of 48 tasks that it created. So it created all these- And boiling it back, let's just remind people, the goal of this project at the end is for you to discover what your site is ranking for in Google Search, but does not-

Yeah ... have corresponding content that it's directing to. So it's missed opportunities for you to- Right ... talk to new people, to- Yeah ... talk to traffic. So this is to create- Yeah, in the SEO- ... and fill in those gaps. Yeah. In the search engine optimization world, this is called a content gap analysis. What gaps are there in our website? What content does not exist that we should be ranking for, that we need to now create because we have looked at the data and we could rank

pretty easily for these keywords and drive more traffic to our site, which means more visitors, which means more conversions, which means more revenue. What do we have to create to fill that content gap? And that's what this whole task is for. And so it's created all these different agents now, each agent having specific skills to make sure that it is a master at accomplishing that task.

And now the only thing left was for me to share those agents with my team, which is very easy to do. You go to the settings, and with Magai, I can literally just hit the share button and make sure that my team is included, and hit Save. And now Precious has access to that agent.

And so we've got our tasks assigned, we've got our agents built. It actually also went through and updated the tasks in Todoist to inform my team like, "Hey, use this agent for this task." Wow. Yeah. So, I think it just made some little details here. Oh, this was one of the things it suggested because there were some stale owner lines. It actually wrote the person who was assigned to it in the description, and it no longer matched some of the tasks.

So it went in, did some cleanup there, and did some more refinement here. Oh, at this point in the conversation, I was getting really tired, and so I was like, "You know what? Send me a Slack message tomorrow morning at 8:00 a.m. to remind me to come review this project," because I was like, "I'm going to sleep, but I don't want to forget to come back to this.

So send me a reminder." And it did because of our scheduled tasks feature. And again, because my Slack is already integrated, it can then schedule this task, and then the next morning at 8:00 a.m. it's going to remind me to come back to this, which was great. And so the next morning it did. I came back to it and continued refining the task and the project. And, of course, the team got started on it right away because they had all the instructions built out for them.

And it worked really great. And so, this content gap analysis, this whole thing that I did from my bed late at night when I was tired, is something that I would, in the past, have charged clients $5,000 to $10,000 for. For that whole process to set it all up and to begin executing it, that was a heavy consulting fee because it was a lot of work.

But- Mm-hmm ... AI was able to do this. I was able to do this in Magai in a few minutes. And you have 47 tasks in it. Are you actually embarking on the project? Are you doing it? Yeah. So the team's already started. There are a lot of tasks that are literally just, they need to go in, copy the task that the AI created, paste it into Magai with the appropriate agent selected, have the

agent write the content for that page, and then they just need to copy and paste it into WordPress. Now, the reason they have to do that is just one small feature that we haven't finished yet in Magai, is the ability to connect MCP servers. We had it in the previous version of Magai. It was kind of clunky, kind of buggy. People didn't quite understand it. So we've been refining the MCP connector so that you can connect literally any MCP.

Any Model Context Protocol-powered app like WordPress has the ability to now connect via MCP to AI agents. Once we build that, hopefully next week it'll be done, or maybe the week after, I can literally connect my WordPress site to Magai and say, "Okay, just go in and make all these pages for us so our team doesn't have to."

So you'll be publishing pages to your website generated by this. Where is the human editor in that process? Is that your- So that's where you want to have some discretion, right? Yeah. You want to make sure that what I will end up doing is I will end up having the AI write a draft of these things inside of WordPress, and my VAs will be the one to go review them and make sure that they're ready, there's no mistakes. So right now, they're going to have to do some prompting, copying and pasting, basically- Mm-hmm

... and then reviewing. Which gives them the opportunity to edit, make sure, and to double-check and to- Right ... approve. Basically an approval process. Right. That's so cool. So when do you think it'll be done? Now, once we have that MCP connector, it'll be no more copy and paste. It'll literally just be, "Here's a link. Go review it. Make sure it's ready. Click publish."

So in the past, this would be about a $10,000 bill, 10 to $15,000, and probably take months to complete. Mm-hmm. At the end, what sort of benefit do you think you're going to see from it? Well, the goal, the hope, obviously, is increased traffic to the website. Yeah. Because we're filling all the gaps with things that Google wants to send us traffic for, or is currently sending us traffic for, but we're not quite optimized to get higher rankings for.

So rank higher in search, which will also rank us higher in AI search or AEO or whatever the buzzword you want to use for it. Because spoiler, most of AEO is just good SEO. So the hope is just increased traffic to the website, which will increase leads, which will increase customers, so forth. Yeah. So just a note to folks, you guys, if you have any questions, make sure you put them in the Q&A chat, the one with the little question mark, because we have quite a few questions. So let's jump to those now.

Yeah. Well, real quickly- Let's- ... I want to give one more really fast example here. Okay. And this is similar to this task, but not quite as elaborate. This is a daily recurring task. My Instagram is now fully automated. And this is an experiment, not something I recommend everybody should do, but it is something that I am personally experimenting with to see whether or not it's worthwhile.

If you look at my Instagram, starting at about right here, the toast man, the guy with the toast head. Every one of these posts was done by Magai. Except for these two. These two I posted manually because they were videos that I generated. But all of these static image posts, I did not touch in terms of brainstorming, in terms of editing the copy, and in terms of posting them to Instagram. This was all done autonomously, fully autonomously by Magai. And it's because of this recurring

task that I created. And you can see, so I didn't even look at today's, so this might be interesting. So every day at 10:00 a.m., my AI is running this task, a very elaborate task that I created with Magai's help. I didn't write all this instruction. I just kind of dictated and gave it my ideas, and it came up with this recurring task instruction.

So every day at 10:00 a.m., it's going to execute this instruction. Essentially, what it's going to do, and I broke it down into simplified steps here. Let me jump into the simplified steps. Okay. So first and foremost, it's going to connect to my source of truth. And at the beginning, that source of truth was a Google sheet.

I wanted it to have record of all the posts I've done in the past so that it doesn't repeat anything. I preloaded it with the last 10. So the first thing it does is it goes and checks the Google sheet. What have we posted in the last 10 days? Okay. Now let's connect to the Instagram account to make sure the connection's there. And now, based on what we've already posted and based on Dustin's preferences here, it's going to...

Oh, wait. Here we go. Come back to this other one real quick. Two seconds. Here we go. Okay. So first it verifies, then it remembers. Says GitHub here because I changed from Google Sheets to GitHub, but just bear with me here. So it reads the history from Google Sheets. It decides on a fresh idea.

Now I told it I want to kind of go between educational content, entertaining content, and thought leadership content. Then it writes the caption in my voice, and the agent that I assigned to it is my viral Instagram post-writing style that I basically gathered from years of what went viral and looking at other viral post-writing styles. It's a custom agent I made for that.

It specifically is not allowed to create fake stories. It's not allowed to invent things that didn't happen in my personal life. No hashtags, except at the end it does add hashtag made with Magai. No em dashes. So it's instructed how to write. Then it generates an image all by itself based on some parameters that I gave it of how I want those images to look.

I wanted them to be cinematic, hyperreal, and with one weird twist. So you'll notice all these images are kind of weird, and that's the scroll-stopping element that I wanted to include. So, a guy with a toaster on his head, a dog with 15 pizzas. Same guy on the table, dummy. All kinds of weird. A gremlin, like a weird-looking gremlin.

So it's instructed to be a little weird. And then what I have it do is I then have it check to see, does this post, before it posts anything, it has to now audit the post that it created I've had it create sort of its own rubric. Does this post seem fresh? Is it honest? Is it scroll-stopping? If it doesn't meet the criteria I gave it, it then fixes it, and it runs that criteria check again until it scores highly on the rubric.

Once it's ready to post, it actually publishes it to Instagram. Then what it does is it goes back to the log, the Google sheet, and it creates a new line and says, "Here's what we posted, here's the date, here's the caption, here's the image description." That way, when it runs again tomorrow, it doesn't repeat anything because it sees the previous history.

And then it also goes back and looks at the number of likes, the number of comments on all the last 10 posts, and then it saves that to the spreadsheet as well, and then it gives me a report of what it did. So connects, makes sure it has the connections, it remembers what we did from previous days, it decides what to talk about next, it writes the caption, it generates the visual, it checks to see if it matches our standards, our rubric, and our quality standards.

If it doesn't, it fixes it until it meets our standards. Once it meets our standards, it publishes to Instagram, then it logs what it did back in Google Sheets, and then it syncs all the like and comment numbers and saves that to the Google sheet, and then it gives me a report of what it just did every single day. Every day, like clockwork. And this, again, is the power of having an agent and having that agent be able to do things, whether one time autonomously or on a recurring basis autonomously.

And it's worked fairly well. I review every single one of these Instagram posts, and there have been some that I've deleted and taken down because I'm like, "Nah, it doesn't match," or, "That isn't good enough." But over time, it has continued to get better because there's one little line in here that I added. Let me find an example here in our weekly shipped summary.

Let me edit this task. So there's a final step called the self-improvement step. So whenever it runs this scheduled task, it is instructed to reevaluate the entire run that it just completed, identify anything that errored, anything that was slower, anything that didn't work right. And then if it found any such issue, silently update this scheduled task instructions to bake in the fix permanently so it never recurs. At the very end, post a one-line self-check note,

and this sends something to my Slack channel. Either no issues, no task changes needed, or a brief summary of what permanently corrected the instructions. So this is a self-correcting, self-improving loop that you can add to every scheduled task so that your agents not only complete the task, but they get better and better at it every single time they do it.

And you are actually publishing and then going back in and doing triage afterwards, determining whether or not it was something that you want. But you could just as easily- Yeah ... put this in to have it published to draft mode, and then actually manually publish it yourself as get a notification in Slack, your post is ready. Right. Yeah. Go for approval. Yeah. Well, I would have to then review it. So the point of this- Yeah ... experiment was, can it run fully autonomously without me having to review it? And that was a success. Now, what I can also do here is if it does produce something that I don't like, I can go in

here to the run. So this was today's run, and I can say, "Ah, I didn't like this image, so from now on, never make images of pottery." I don't know. Yeah. And edit that task, and from here on out, it will make sure it doesn't do that anymore. So you can not only just send agents to do tasks, you can also schedule them to do tasks on a recurring basis, and you can have them self-improve, or you can give them feedback on how to improve. Creating essentially your own AI employee workforce that only gets better with time.

And certainly, for things like generating sales reports, this sort of stuff, this kind of multi-step, quick evaluation analysis is going to be awesome. Cool. I know that was a lot, but this is a big deal. All of this is really compounded.

A lot of this is so important that all of it sort of, you understand how it all works together, that I kind of wanted to just fire hose all of it, and we can kind of break down as much as we needed via the comments here or- Okay. So let's bang through some questions. We've got about 10 minutes, we'd like to kind of keep it to, so we'll get through as many of your questions as we can, folks.

If you've been posting any questions in the regular chat, make sure that you repost them in the questions chat. So Michael's saying, "It sounds like, with skills, your agent can do sequence actions," which we just showed. Yep. "What tips do you have for agents to do multi-step tasks efficiently?" Well, if you know what those multi-step tasks are, like you actually know you have predetermined the steps that you wanted to take, just be clear about it.

That's really all there is to it. Yeah. So what happens, Brad's saying, "What happens with personas? Do we just delete them now?" No. So Brad, personas are agents, agents are personas. It was literally just a name change. Yeah. But like I said at the beginning, it's not just a name change You now can do more with them. So your personas are still there, they're just called agents, and now you have the ability to add more skills to them to make them more elaborate, more sophisticated.

And Brad was also asking a few things about... Okay, let me, there we go. Does that come up now? There we go. He was asking about personas. His first question was where workspaces fit into this change. We've already talked about how these agents work within workspaces, which it gives them the context. But if you've given more detail into the agent, which has priority?

If you're using workspaces properly, it's not about priority. Workspace is for organizational context, right? So if you're using it for the right purposes, like this is the project and this is our business, is really the best example for using a workspace or a project.

It's for business details, not task-related details. You shouldn't be putting brand voice into your workspace. That's the role of an agent. You should be putting static, unchanging information in the workspace. Pricing information. Terms of services, product.

Things that aren't hierarchical in nature. They're just, this is the truth. Yeah. So the priority doesn't really come into play there if you're using it properly. So Michael's asking, should we be updating our legacy personas, now agents, with skills that we let the system define and develop? I would explore it. If your personas are working really well now, maybe try augmenting them with some skills and see if it helps.

Because again, this allows more sophisticated... I keep defaulting to persona. I have to catch myself. These are just allowing more sophisticated instructions, more elaborate agents than were previously capable. So if the current one is working fine, then you don't need to change it. Like for example, I have several agents that I haven't added skills to because they're not necessary.

The first level instruction is all it really needs to do its job. But if you want to condense some of your agents into one all-powerful agent, that is one way to be more efficient. Marketing example. You don't need a social media marketer agent and then a social media ad agent. You can create one agent that has all those skills in it, and it will appropriately choose the right path every time.

Yvonne's got a long question. "For my business decisions, I want to use multiple agents." Remember back in the day, similar to when we learned about the board of directors, which you showed us about four months ago. Mm-hmm. Example from me. Step one, use selected agents to build my learning model. Step two, taking the completed outcome and having the selected business agents use the outcome, talk to each other to discuss and agree on the best path forward. How do I add agents separately? Depending on the business task, I want to select different agents.

Well, you can't have multiple agents working in the same response, but you can at any time, just like you always have, you can switch agents in between responses. So let's see. Start a new one. So if I want to start with...

Do I have Groot? I do have Groot. So if I want to start with my Groot agent, and we'll switch to, let's just choose Google flashlight because it's just a simple demo. So I can start with Groot, have Groot respond, and then I can switch to my assistant, and I can say, "What's the date?"

And the assistant will then take over. So you can switch between responses, but you just can't have multiple agents working in the same response. Okay. All right. Steven's asking for an update on when we can add custom MCPs.

Like the MCP. It looks like the MCP- Yeah, we were hoping last week it'd be ready, but we just found a few quirks that just weren't ready for production yet. Next week or the week after is really looking like what it's going to be. Provided that it all works the way it should. So with agents that come with Magai, how can we see what the skills are that you have given them?

Right now, there's no way. So if we share an agent publicly, there's no way to view their skills. We're still trying to figure out the best execution of this because what we eventually want to create is an agent marketplace. Just like we had a persona marketplace before in the previous version, we think that we want to revisit that idea where creators can share their agents. Now, you can do that.

You can currently share your agent publicly with anyone. All you have to do is go to the settings, scroll down, and there is a public access toggle. You can turn that on, save your changes, and then you'll have a public link. So you can copy that link and give it to anyone with a Magai account, and they can add your agent. Now, the thing with this is a lot of people creating agents do not want people to be able to see their skills because maybe there is some proprietary data.

Maybe there are some proprietary things that they're doing with those skills or with those instructions, and they want those guarded And rightfully so. If they're using proprietary information and data to create a really high-level agent, they should be able to maintain the integrity of their trade secrets or whatever's involved in that. So for that reason, we're not displaying skills.

We're not displaying instructions on shared agents. But as always, we will do our best to open source our own agents. We have currently a Notion page that has all of our public personas listed in there. But as we continue to work on building out the new agent marketplace, we'll have an option where some agents will be open source, which means you can see everything.

Some agents will be closed source, where you can't see everything. So that's just in progress. One of our newcomers, Steve, is asking, "Can I ask an agent to build a website as well? I'm relatively new to this. Is this a stupid question?" No, it's not a stupid question at all. No. Last AI Academy, was it the last one, Steve?

Or maybe the one before we talked about- The one before, I think. Yeah ... Yeah. One before. If you go to magai.pro/aiacademy is the short URL. You can go to Academy Replays, and it was number 88. Yeah, it was the last one, Publishing Web Pages. Yeah.

Or maybe that was the one you did yourself. Was that yours? No, this was one month ago. Yeah. We did it. And it's something that we'll return to, but there's a couple of ways to approach that. You can actually build web pages, right, using any of the AI agents or through the MCPs. If you have a web page creation tool, you can also interface that way. But yeah- That's a whole lesson in and of itself ... the development of websites is...

Sorry. Go ahead. I was going to say, that's a whole lesson in and of itself. Best to start with AI Academy number 88. Yeah. That'll give you a start, for sure. But it is very doable and a great work case. Okay. Sue's calling you out. "I'm worried about the number of times Dustin says, 'I didn't read all this,' with a sophisticated user who understands his business. Well, something might squeak past that, but most users are not at this level." Sure.

Yeah. I agree, Sue. In all of our AI Academies, you'll hear me constantly say, "I didn't read all this" or, "I'm going to skip through this." Mostly that's for demonstration purposes because I don't want to read through it for the demonstration. But one thing I did highlight is I scheduled that task for me to review the next morning. Remember, I scheduled that reminder for myself to go revisit the project. That's when I went back and reviewed everything.

And I think we can give a little bit of a disclaimer here. We're looking at the capabilities of using AI. You have to choose yourself what your thresholds are and how you're going to apply the AI tools. And there's going to be lots of cases where we show a capability, which is something that we might not choose to use in the wild, but we still need to understand what it does. And coming up with examples that are going to be absolutely clean and best use cases as far as best practices, that's very difficult for us to do.

So basically, I think what you have to do is you have to look at the examples that we give and then determine how the processes fit for you, not necessarily the outcomes. Yeah. Dave? And the other thing that you have to understand, too, is I was using Claude Opus, and this is something I know about Claude Opus. Unless you specifically tell it, it's going to be very overly verbose. It likes to talk.

And I should have refined it from the beginning, just saying, "Be concise," and it would've been more concise. But that's something I was highly aware of as I was doing this- Brother ... late at night and from my bed. I'm going to wrap things up in just a second. We'll stick around for a few extra minutes, but we're going to let people go who have to go because we've reached the top of the hour. Heather is asking, "Is it possible that the content from July 29th is captured another way?" We had a disaster. For some reason, my computer would not share my screens, and so I wasn't able to show what I was planning in the July 29th, so I never reposted the video.

So it's not available, and it won't be available because it was pretty boring. It wasn't good at all. And with that, I think we're going to put a button on it. We'll wrap things up fairly quickly today. Dustin, fascinating stuff. I think one of the things that we should touch on next week is we should touch on how do you go about finding new agents that you can then incorporate in your life.

Searching for them and then uploading them might be something that would be really valuable for people. And we can go from there. With that, thank you so much. Stay cool. Doing my best. To each and every one of you, I hope that you have a great week, and we will put a button in it. We will see you all in two weeks, sorry, in two weeks on the AI Academy. And then once we hit September, we'll be back to weekly. All right.

Till then, have fun storming the castle.