AI in Context, Ep 01: Delegating to agents

Machine transcript of the replay, unedited. Times link to the moment on YouTube.

0:00Well, let me get this started. Like, okay, I've got Abhinav as well. Abhinav is an awesome, another developer. My, like, the goal in this is I want to do this as a weekly thing, right? It's really, there's so much to talk about when it comes to AI. And what, like, the biggest advantage I have with regards to this, I've realized, is that I've spent a shit ton of time on this. Like I'm hitting my 9 ,000 now. I want to get to 10 ,000 hours by the end of the year. I'm slated to hit my 10 ,000 hours with agents from in about end of November, beginning of December at this point, based on the 12 hours, 13 hours a day I put into these models. And more or less, I don't do anything on a computer now without going through a model.

0:49Like every single action I do on a computer goes through an AI agent to some extent. Like I have not filled a form on the web in three months. I had some, so in Canada, when you do digging on your property, you need to send out a message to like the authorities. So they come like notify you of where the utilities are, right? And it's a really complicated form because you have to have like the map and you have to like draw out where it is. I had Codex do it for me. Like it's, there's just so much stuff like busy work that you can do today with like existing models. And my problem is like a lot of people who are talking about AI are focusing way too much on tools, on like Claude is this and ChatGPT is that and you can do this. And it's like most of you have already got enough experience with this to realize that it's not about the tooling. It's really about the methodology. It's really about the like the mindset that you approach to this. So my objective with AI in context is really about putting AI in the real life context. And a lot of you I've talked to individually about this. It's just like my views on AI is very much that it is a very big lever. It's an amplifier. It allows you to do like 100x more.

2:04But it's just how do you know what to delegate? And how do I delegate effectively to agents in a way that I'm not lost? So this is like first part of a series. The goal is to like kind of do this collaboratively with you. It's not like I have a game plan. And so what I did is I have decided for this one to do a, so a, a, a, like a very much a, here, let me put myself on the screen so you guys can actually see me while I'm doing this. There you go. So like my goal is I've got this, I've got a blank ChatGPT instance, right? I've got really nothing on it. So we're going to build this together over time so that it's like, I don't want any barriers to entry because I've put in several months and a lot of my money into building my own system, but that's not attainable or realistic for a lot of people. So my goal here was how can I show someone from scratch so that I don't assume anything about your process. um i will if this is not like a course i want you guys to interrupt me i want you guys to ask questions you guys are all really awesome people so like you guys have ideas and you have experience i will cut you off if you're like trying to to give too much advice in that because like we can have that q a you can talk about it after i'm going to take like 25 minutes to go through what my thoughts were and then i will be able to have that kind of conversation so let's get straight into it. Like, here's the baseline.

3:41It's like a lot of people have this idea about, I use Gemini, I use Chat, I use Claude, I use GrokBot, I use all of these new tools, I use OpenClaw. And it's like, I don't care. It doesn't really matter. At this point, models are, they're all very, very good. Like if you're an open source kind of person, and you don't want, you're like, you're really big on data sovereignty, and you don't want anybody, go through OpenRouter, go use like a a ZDR, so like a zero data retention, go use GLM 5 .3. It's just as efficient and you can actually do everything I do with opens like relatively low cost open source models. It's going to cost you a few hundred dollars per month if you're using it at my level. And that's like, there's no question now about the tools. And this is not, this is going to keep getting worse. What actually matters a lot more I've realized is this, It's this concept of when you're doing models, when you're doing actual work, there's a question of defining what is it I'm trying to do? What is the outcome I want to achieve? And what is the evaluation I'm going to use to determine if I've achieved that outcome? And then finally, what is the frequency at which I want this thing to be done?

5:03And it sounds like a lot of busy work, but in essence, it's more about guiding your attention with regards to getting shit done. It's very easy to feel busy and it's very easy to send the model off to do something for you. But it's not actually realistic or in your interest to send it in a direction for doing real work. I'm talking like if you're a developer wanting to really add a feature. If you are a business, like I'll use Neilda as an example, because like Neilda was talking to me about I want to be more active on social media. Right. And then it's like, OK, great. Let's let's break that down into subtasks. Right. So I want to be active on social media. What does that mean? Well, it means I want to post more on LinkedIn. It means I want to be more active in my discord.

5:58It means that I want to look at podcast opportunities because I want to have more distribution. It's like there are real things that you can break down in terms of like, okay, I want more social media presence. Well, these are the actions I can take with regards to that. right so if i use this and i say okay let's let's choose look at podcast opportunities right here as the task well the outcome in my in my mind would be book a podcast opportunity and then it's a question of like well what do i need to do with regards to that well in order to book a podcast opportunity i'm going to need to reach out to the people the podcasters so i'm going to need to find them, I'm going to need to send them an email, and then I'm going to need to follow up. Like that's kind of, if I were to break down that task, that specific task into its subcomponents, that's kind of what it looks like. And the eval is, in this context, would be something like, is the system sending out X emails per day, right?

7:16The frequency then would be on a daily basis. And it's just like, this is actually an agent. There's no bells, there's no whistles, there's no nothing clear. This is just like a model loop that has a task, that has a process, that has an outcome, that has a way of determining whether or not it has achieved that outcome. And then it wakes up on a certain frequency and does those things. And then that's it. And so people are, when I present this to a lot of people, especially like businesses that pay me to do this for them, they're like, well, that's really simple. And I'm like, yes, it really is. Agents are made to sound like they're really complicated. They aren't. But what makes agents complicated is A, getting good at delegating and actually defining the things I just did, and then starting to manage it when there are multiples, right?

8:21It's one of those things where the compound effect is really the name of the game. Doing a single routine for yourself or a single agent is not that impactful. But when you launch an activity back in June and you set an agent on it and then you completely move on with your life, the agent doesn't forget. The agent keeps doing the task that you had told it to do. This leads to like the developers of the room like Eric and David are all thinking they're like, yeah, but, you know, agents over long time horizons could get funky and et cetera. Yeah, let's put that aside for now. I think that's like that's a that's a future problem. That's not a present problem. I think when you're starting off with agents, there's really this concept of the building blocks of starting with agents is much more simple than you'd think. It's I have a task I want to achieve. I need to define what that outcome looks like, what that process looks like. How do I know if it's good or not? And what frequency does it need to happen?

9:26What that leads to for me is this. so last last month in the last 30 days i spent twelve thousand nine hundred ninety three dollars on api credits um why because astra is bloody expensive and i have stopped using astra since soul came out but all that to say like you've got this this is about i've got 720 agent flows happening right now for my business they're very some are really small things some are bigger things. The podcast thing I have is actually working. I've booked three podcasts. I've got my first podcast coming out in a week from now. I recorded it about two weeks ago with the person. So like it's the compound effect, right? I set this up maybe a month and a half ago. It took three weeks to get my first podcast. And then, well, I have one. Now I have two more in the ropes. And it's just like, you do these things progressively and adding on bits and pieces to your business and then it starts actually producing stuff so are there questions before i move on to showing you guys how how to do this well the question is how do we do this all right part two How do you do this?

10:48So my goal here is how do I do this? So I'm actually gonna I'm gonna go ahead and what I'm gonna do is I'm going to take this right here task outcome eval frequency I'm gonna I'm gonna write that so before I go forwards because I know that not everyone's at the same level with regards to I'm gonna use ChatGPT for this example I think next session I'm gonna use Claude, and I'm just going to switch between them because the goal is not the tool. If you ask my personal opinion, I believe that ChatGPT's app, the Codex app, is literally the best harness right now for general purpose AI. It's a very, very good tool. I think Claude's app is also very good and very capable. There's not really a major difference. The tool doesn't make the difference is what I'm trying to make clear here. That being said, I'm going to be using ChatGPT for this one.

11:44So first things first is there's a lot of stuff in ChatGPT that people don't know about. And I'm just going to go over them really quickly. I don't want to do a tutorial on ChatGPT, but I do want to give you guys some pointer because their product is good, but their product design is sometimes a bit wonky and not very clear. There's a lot of toggles in ChatGPT. And if you want to start like managing agents for a professional business setting, the things I recommend for you are up here, you have a dial chat and work. Chat is literally it's your Google search. That's that's the perception I have of it. It's very much if I have just like a random question to ask, I'm going to use chat for professional tasks. Oftentimes, it's not the right it's not the right call. Work is better. Work is actually much more of an agent harness in the sense that it has access to connectors, it can use projects, it can actually create files, plugins, etc. And it can, you know, do more. That being said, neither of them are what you want with regards to actually doing this for your business. I don't know why they haven't done this. I think they're going to get rid of it. But here, if you click on ChatGPT, and you press this little button, you actually have Codex, which is another version of ChatGPT for more developer crowds, but is actually just a better overall version of the tool.

13:11And it's always, so ChatGPT, Codex is always on, it's always on like dev mode. It's always leading to, in this case, a cloud environment, and I'm going to be able to choose an environment. I'm going to be able to create an environment, connect GitHub, et cetera. So when you're doing agents, it's really important that you use Codex and you use GitHub. Why? Because you don't want all of your context to live inside ChatGPT in these temporary chats. You need a way to show yourself as well as others what have changes happened over time. GitHub seems really complicated. It's really, it's just a ledger. It's just a place to say, I did this thing on this date. And then when I do another thing on that date, I'm going to add it. So you can go ahead and connect it to GitHub. I'm going to go ahead and connect it myself.

14:08Go ahead. I'm going to use my 1Password right here. There you go. And then I'm going to be going back here. And I think I need to. Good. We are good. There you go. So I have all my repositories now, and I can create an environment. I can go ahead and create something. I can go ahead and get something. I'm going to take a random one that I haven't used in a very long time. So I'm just going to use, I don't know, Gatsby themes, which I haven't touched in like six years. And this is like the baseline environment setup. There is no more setup. You are now actively creating an agent. So as this is happening, so what I'm going to go ahead and do is I'm going to wait for this to set up.

15:06And then I'm going to use my podcast example as the example, right? And what I'm going to do here is I'm going to say, I want to set up a routine that is going to find podcasters in the business space. If you do not have open source accesses, let me know so that I can provide you with a scraper that will help us do this. Following that, our objective is going to be to find podcasters on which to promote our business, to talk about our experience, and be able to create some inbound on the topic. Note here that the GitHub repo I have given you is not at all related to this task. We're just using this as a staging crowd. We're going to delete everything on this repo and start from scratch.

16:02So, there you go. This is how I start most of my prompts. I start them because I've learned over time that prompting is, I didn't get more benefit from spending a lot of time working on the initial prompt. I got a lot more benefit by working on it iteratively with the models, especially as they've got it better, right? So my recommendation is start with a vague idea and work with the models in order to fine tune that idea. because as a reminder, if I go back to this thing, my objective is really just to come up with a task, with the outcome, with the eval, and with the frequency. The rest of it, we're going to fine tune over time. Like agents are not a set and forget kind of thing. You create them, then you keep a watch on them for a little bit for normally it's a few days. And then once they're working like you want, then you let them go for a week and you come back and you check them, right? So like right here. Yes. I have a quick question. So are you, um, when you connect on GitHub, you have your repo, are you assigning one agent to a repo or all your agents in a single repo?

17:19um i found that it's better to have a single repo for your business i i tried multiple things so the separation i make is uh so i've got this actually i can go show you that if i go to my audit tool so i'm just going to bring it into this browser and i go by so if i go here and i go to my usage you can actually see here that my lemon repo is my main repo it's my company's repository and so a lot of my agents are running within my lemon repo i've also got other like marketing i've got other clients works that is smaller you see that actually most of my clients are all held here so i'm dealing with about 16 companies right now in terms of my business and how like the clients I'm serving, all of those clients represent less than 30 % of my spend in terms of compute costs. And all of these costs I actually forward to them. So you see here, I know my exact costs, so I can just forward those. But to answer your question, I recommend having a single repository for your business in your life and then creating maybe secondary repos. So I did one for marketing. I wanted marketing to be separate from my business because there's sensitive information like my taxes and my signature documents and things like that that I don't want necessarily being in a repository that is going to be used for outwards tasks like marketing related stuff.

18:59That being said, when you are on your environment, you can edit the environment to have multiple repositories. Like especially when you're on the local ChatGPT or local Claude, you can give it access to the local folders that have both of those repos, right? So you don't need to like you constantly think about this. What I recommend is create like a general purpose environment that has access to all of the repositories that are part of your day -to -day and let the model through a CLAUDE.md delegate to the right repository. I've not had any problems with that so far, so that's what I would recommend. So now that we're here, like it's saying, okay, Lemonbrand is my business, is my business.

19:50Go check out lemonbrand .io. So this is one of the big things about agents is I don't try to over -engineer them. I don't try to overthink them. I know that the first version is going to be the worst version of the agent. It's only over time, over a few iterations, that I'm going to get to something I'm happy with. And so I've abandoned, like I have a more laissez -faire attitude about trying to get it right, right from the get -go. I'd rather get it right over iterations because I don't know what I don't know. And when you build agents, that's constantly the situation. So I'm trying something new. Otherwise, I wouldn't be using an agent for it. I would have already done it. And so because I'm trying something new, I have to assume I'm going to get it wrong. Because I'm going to assume I'm going to get it wrong, I'm going to need to do iterations on it.

20:41So what should the first version of the routine deliver? So this is, I like the fact that it asks you questions, but again, repeatable script that produces, right? Scheduled prospecting routine with new leads. I like scheduled prospecting routines because, so this comes back to the second thing and developers are gonna think about this, is you have kind of two choices with regards to agents. Like the easy path is managed agents. So this is Claude routines. So Claude routines, codex routines, all of these things, all of these scheduled things that happened within the application. So if I go on to my actual application of ChatGPT, I believe I closed it because it eats up my memory like it's nobody's business.

21:43And I go open that up just quickly to get you guys a screenshot. I'll show you that like there are routines that exist and what that looks like in reality is very simple within their program. So as soon as that opens up, it's right here. And I'll be able to set that up for you. My goodness. Anyone else have problems with the fact that Codex just eats up your RAM? Or is it just me? Senior eats some stuff sometimes, definitely. Right? Especially starting to use it. Oh, good. Okay. So I'm not alone. I'm glad. Here's what that looks like. So here, I've got this one. First, perfect. So this is what a codex routine looks like.

22:38And then I will show you the schedules right here. Schedule like this. There you go. And get rid of this program before it just makes me die. Quit. Perfect. So this is kind of what it looks like within, within like my actual, this is verified. No, this is just like a website that I built with one of my co -founders and just like having fun with this, like fully agentic business. But this is, it's got two scheduled tasks around to it. I've got a daytime routine and a nighttime routine for it because like I I want to conserve tokens. So I don't want to be always like during the daytime, I need to be on top of stuff. So it's running every 10 minutes during the night. I do it every hour. So it's just like, it's just about token conservation there. And then I've got the task name. I've got the actual prompt that was built with codex. I've got the schedule, et cetera. This is what managed agents look like. Like we have this idea of this agent where you like see it working and and you've got this fancy dashboard and everything, I'm sure we'll get there eventually, but we're not there yet. I think if you wanna be more practical and actually do stuff with agents, this is what that looks like in production.

24:00Now, the issues I've had with the easy path is that there's two issues. The first is sovereignty, which means for most people this doesn't matter, is you don't own your data when you're doing this with Codex or ChatGPT within their application. This is not a problem for your business. It does become a problem if you want to sell this service to some other people. Now, you have two options with regards to setting up agents in other people's businesses. Option one is to do it in their ChatGPT that is under their licenses and in their organization, which is a very good option. Option B is to do it the hard way. And I'll go over the hard way or the harder path after. The second thing that's like a negative is that this is either cloud or local. So if you, a lot of the times, these scheduled runs are happening on your computer. That's not a problem if you're Jeroen and you have a Mac mini that you just use and have given to your agent.

25:13which I love, by the way. But you do need to give it a computer to always be running on if you're going to be using this for production. And that's a pro and a con. It makes it very easy to spin up an agent. It's easy for it to go down at some point. And if you're on a holiday for a week, and there's a power outage, and it doesn't come back online, well, you're screwed. So it's kind of the things that you need to think about with regards to managed agents. The hard path is, and I say hard path, but it's more owning the infrastructure. So what these companies are actually doing is relatively simple, is they have virtual private servers, so VPSs. I personally like using DigitalOcean for this. You can use AWS.

26:06You can use Hetzner, how it serves up. You guys know what I'm talking about. The devs know what I'm talking about. You can use like a lot of managed services for VPSs and you can have your agents running on this with a SDK. So a software development kit. So this is option B. I'm not going to explore this. If you guys want me to explore this at some point in the future, this is a much more in -depth conversation that goes into the whole development cycle and how to manage these and how to do everything. I don't want to explore that right now. I want to get back to how do I do this for myself, right? So I've done this. How often should the routine run? I want it to be weekly. I want it to be emailing a short -ranked shortlist. I want it daily.

27:01Let's do as basic as possible. I don't want any connectors. I just want a weekly saving a rank CSV in the workspace. And you see, this is all there is to it. And it sounds ridiculous to say, but like, all I need to do at this point is I've given it the task, which is I want to set up a routine. Now what I'm going to add on to this is, I also want to set up an eval set that determines whether or not the agent is doing its job. The outcome here is for it to be successfully finding new podcasters and not repeating the same podcasters every week. And eventually the eval is going to be a question of when I send emails to these people, will they respond? Now in the context of this demo, I want us to set up the eval being are we finding new people? and then I also want us to figure out exactly which target we're going to be defining kind of our ICP for the podcasts.

28:15So while this is running, any other questions coming up? Are there things that you would like like that are more relevant to your business that you would like me to create a new chat right now and explore with you or just maybe express my thoughts about how to delegate to agents and create these? these routine agents. So you mentioned right now this agent will exist on OpenAI cloud -based, right? And what about all the other plugins, like to your emails, for it to send emails, would it automatically have that based on whatever you've configured with OpenAI? Exactly. Exactly. So you see here within the environment settings of your cloud on OpenAI platforms, you actually have environment variables. So if you do have API keys, if you do have services that you want to give it, you can actually go ahead and do it through this. You can also give it access to your connectors through this. So Gmail connectors and drive connectors and all of these things. And so, for example, when you're doing graphic design work, when you want to create assets on a cron job every single day, you want a new asset bank generated for you to do posts on LinkedIn and social, et cetera.

29:34Well, this is a great place to connect your drive so that when the agent is producing a new asset, it's part of the skill file to upload it to the drive. so you have a human accessible place to go see all the stuff that the agent creates. Cool. Any other questions relevance to you? I want you guys to talk about yourselves. I want you to share your problems so I can answer them right now. This is the map. Sorry. Go ahead. I was just going to say this isn't a question, but I noticed a lot of PRs in the list and that's a really good practice. I was happy to see it.

30:24Oh, you're talking about all of my PRs right here. Yes, a lot of people don't do that. That's a good way to keep things organized. So yeah, I'm going to emphasize this. Is that a pull request? Yes. So what is a pull? Sorry, I get ahead of myself. What is a PR? PR is a pull request. Now to translate this into non -dev speak, it's a change to the thing I made. so like i i start at number one and then afterwards it goes up to number two every time i make a change it increases this number now agents don't care if it's pr number 2600 humans it's a lot it's it's too much for us agents don't care and it's it's it's preferable for you to over commit and and like constantly make changes and have like this timeline of you evolving. I have a day on my GitHub, I believe, that I did, I think, 400 or 500 pull requests, and four to 500 changes in a single day. So yeah, there's no, I agree with Eric, doing a lot of PRs is a very good practice.

31:39Do you review the work yourself? I use a series of tools such as CodeRabbit as well as Claude Review to go through. I have found that since the Astra fable moment about two, three months ago, code quality is no longer as big of an issue. unless you're working on a team with like high, where there's a risk of something really going wrong and impacting a human, if you're just doing your own projects, no. There's like even my client projects, I trust the AIs as long as I've set in motion the right, that I have a clear idea of what I want. That's really the determining factor.

32:29Before I go, Atharva, you were going to say something. Could you, do you wanna share? yeah um i was just asking because you had earlier you had said that the first run from the agent is always quite poor and so you like kind of want to iterate with it um have you tried like uh agents empty is that something you would recommend like if you already preload that context into the agent does it really help because then that way i guess maybe you don't have to iterate as much and it like streamlines it? Is that something you use in your workflow? So yes and no. So to everyone wondering, agents .md or claude .md, although both are now standard and it doesn't really matter, what they are is the way it works when you're chatting with an agent is you have here a user message. you send this to the agent and then the agent receives this, right? And the agent replies is like agent replies.

33:41What CLAUDE.md does is it actually, before every single term, it will actually insert your agents .md. That's what the agents .md represents. That's what it does in terms of the flow. So it's like here and then it's going to do the same thing. That's kind of what AGENTS.md represents and that's what it does. Now, the issue I have with starting with an AGENTS.md is I usually don't know how it's going to break. And AGENTS.md is a very good place to put information about how it breaks. So that the, because like models, every time you restart your instance, every time there's like, So more or less every 12 hours, because they're trying to preserve their compute, every 12 hours, whatever's happened is going to disappear.

34:41And then there's going to be a new agent that's going to have a new agent is going to appear. And they're going to have like the past history together. And they're going to be like a new person. This is where AGENTS.md are very useful. in the context of making a new automation from scratch, I don't know where it's going to break yet. So I use the AGENTS.md as the tool to keep it on track over time. Now I'm going to make a caveat here. We were talking a bit earlier about the fact of having like my organization repo. And then within that repo, I probably have multiple, like you know I have project one, project one, project two, etc. Now AGENTS.md are very useful in this context because your agents MD can actually list out the projects and their peculiarities. So in the context of working day -to -day with models and understanding how that works, their Our agents are agents .md files are extremely useful, but in the context of creating a singular, like here, if I have this as agent one, so this is a routine that runs on a schedule, it's not necessary for me to have the agents .md from the get -go. Eventually, I'm going to want to create a specific agents .md for this project that indicates the problems with it and the things that it needs to be. I call these like blood taxes where you did a thing, you figured out that when this happens, this breaks, so I'm going to now figure it out.

36:29That's where you'd add it, but not before. Does that answer your question? Yeah, yeah. Thank you. Wonderful. I'm trying to go and break down the concepts as much as possible because I don't want this to be like inaccessible to people. So at this point, we're almost done here. So summarize what we've done. Activate the routine schedule. And let me know what's missing. So this is the kind of stuff. One of the things with agents is because we're all there multitasking all the time, it's very difficult to track what this conversation was about. One of the habits I've gotten into is whenever I'm doing something else, I'm talking to someone, I'm doing a call with someone, I'm working on a project and I task switch. I need to constantly spend some time asking the agent to summarize for me what we've done so far because it's not worth my time to go up and scroll up and understand exactly what's been done. the agent is much better at summarizing it for me than anything I could do.

37:47So in this context, it says workspace schedule is active and verified. It runs Monday at 14. Next run is October 12th at 10 a .m. Yes. Can I ask a question? I can remember when half a year ago in the times that PRDs were important and you spent like a gazillion iterations before you started anything. which I loathe. You always used to be really proud of the fact or really outspoken about the fact that it was really important to read everything. You would never let the agent do anything without reading everything. Here, it seems to me like your trust in the agents has changed. Therefore, a summary is good enough for you now.

38:43Is that true? It is true. Okay. Well, I just wanted to double check. It's an excellent point. In my case, laziness took over, but in your case, the trust of the agent. There was a point this year. For me, it was around August, July, August, where models got smart enough. Very much like since I would not do this with GPT 5 .5 or Opus 4 .7 or 4 .8 or even Opus 5. I would not do it with those models. The thing is, I personally use Fable, Opus 5 .5, Astra, Sol 6 .1, GLM 5 .3, DeepSeek 4 .1. More or less, if we go together now to the Intelligence Index, which is a great evaluation set by artificial analysis, and you look here at everything here in terms of all of the like the the the anthropic models gpt6 astra gemini i'm going to ignore for now because like their harnesses suck um all of these down even to grok i agree grok would be that these i don't use as much all of these models are sufficiently intelligent that i don't feel the need to review specifics as long as i i have a good idea of exactly what the outcome in the eval looks like. So what I'm trying to say here is summarize what we've done, activate the routine, and let me know what's missing. The last thing I would ask is how can you prove to me that this will work?

40:37Can we do a dry run? this will act as the basis for our eval. So there's a question now at the point where agents are, because I'm using here 6 .1 sol light. Why? Because it's less expensive. I can really tell the model to prove to me what it's doing is right. And the eval set is what I'm going to be spending a bit more time on. So this is particularly important when it's dealing with customer support replies or things like that. It's less important when it's like just like a random, I want you to check the internet out and find articles for me because that's just sending me links. But I'm going to spend more time on the eval side and less time on reading everything. And yeah, we're past the point where PRDs aren't that important, I'll be honest. I think models have gotten very smart.

41:35Question. I noticed that you are using Astra Light and on the website that you shared artificial analysis, a lot of the examples are, let's see, Astra Max. Yeah. Should we go for the max or is light enough for what we want to do? I would say that in 99 % of use cases, the model family takes precedence over the reasoning level. So what I'd say is, if you can afford doing more, if you can afford, and I mean financially afford it, I would use the middle of the tier, low tier. So if I go back to here and I look at Sol Lite, I have all of the models over here, and I look at Astra. Astra has light, it has medium, it has high, it has extra high, and it has ultra. If you do ultra, you're going to use up your weekly quota in probably 20 minutes, if not less.

42:42I would do Astra if you are in a very, very, very difficult situation with a problem that would impact people in a very critical business system. That is literally the only use case I would use that. And personally, I almost never do. I seldom go past medium in terms of reasoning levels, because as I said previously, I'd rather figure things out through evals and through real effort than doing it, like spending more money initially to get like a better version. I think time is a much better judge than a better reasoning model. I appreciate that. Thanks. Cool. No problem. Hey, Simon. Yeah. David here. Hi, everyone. Yeah.

43:36Quick question. How, in terms of workflow, how would you plug a new agent routine system on an existing routine? For instance, if you think about marketing, you're going to want to think about strategy, updates, and being up to speed with everything happening. and in a very high -speed environment you're going to think about uh content production lead generation etc etc so and that leads to my second question which is very practical in terms of GPT Codex sessions or Claude Code sessions would you privilege um working in one single session for an entire business unit, marketing, accounting, et cetera? Or would you start from scratch again on every session, knowing that sessions don't communicate well within each other? Yeah, really good question. So two things here. The first is if you go here and And if I go back to ChatGPT, I believe here, you can absolutely use the projects function. The reason I'm doing this online, I strongly recommend you download the ChatGPT app, which is a much better user experience. But you are, same thing for Claude, you can set up projects and you can do things in projects. And I do recommend you set up projects for yourself that are related to each business unit. it.

45:13Just it's not actually for the agents, it's for you as a human, because it's very difficult to keep track of everything. Now, to your second, your first question, which is, I think, more important. So you gave the example of marketing. So I have marketing as a function here. Now, do I, if I summarize your question is, do I add on functions to an existing agent, if I'm going to update it to add more features or do I do something else? So I'm going to break it down and say like I have a LinkedIn content posting agent. Now this LinkedIn content posting agent is not writing the posts for me, but what it is doing is it has inside of it my Claude or my AGENTS.md And within that AGENTS.md, it has my voice examples. It has things like the type of my voice examples. It has successful posts. It has themes. It has opinions that I held, et cetera. So that's in my AGENTS.md. And now its job is, so if I go back up here to my thing, its task is produce one draft per day of a LinkedIn post based on a theme within my quad, my agents .md.

47:02agents .md. So this is its task. Its frequency is one per day, etc. So this is just doing the text. Now, some people get fancy and they want to add images. They want to create much more stuff. What I would do is actually I would create a cascade, which is this runs first, this LinkedIn content posting or you know I could add another level here and I will do a LinkedIn research agent and so this the job of this one is very similar to the podcast scrape new topics so I would give it a scraper so I personally use Apify for this then I would give it a list of relevant channels slash usernames and it would go find a raw material and this would be to its task would be to produce one 500 word document with what's recent news relevant to our topics and its frequency would also be once a day. And this would actually be the trigger.

48:18This agent would actually trigger this agent. And so once this agent would be done, this one would then trigger a new agent, which would be the LinkedIn asset creator agent. So to answer your question, you should not be adding on to an agent. If you're adding on multiple tasks and multiple outcomes to an agent, it's because your agent is too complex. It needs to be broken down into smaller, more bite -sized pieces because like, if I do everything in the same agent, how the heck am I going to know what's broken? Is it the content? Is it the research? Is it the asset generation? I don't know. Each of these are a lever that I need to be able to fine tune on it by itself. And so you generate new tasks and this can be applied. So if I think I'm doing your CPA, right? If I'm doing bank conciliation, right? Well, my bank reconciliation is I've got to have a, I've got my bank statements that come in, right?

49:18Now, the thing that's going to be happening here, so if I go to my task, it's go through the PDFs, extract the line items, and input them into a database spreadsheet, right? So the bank statements come in, they go through the PDFs, they extract the line items, and the output is validate against the PDFs that there are no errors in the database. And then the next step after the bank reconciliation would be another agent. and that another agent would be looking at inputting these into my accounting software so you see i would where does the triggering and the multi -agent where does that sit within codex yeah or do you use another layer that was my next question so i have to trigger all of that so thankfully actually managed agents can communicate to each other this is actually the topic of next week's webinar. How to create cascading chains of agents is what I was going to cover next week because it's already been an hour, guys. But yeah, this shit is super interesting. But yeah, so you can actually go ahead and ask. I do recommend is how do I create cascading agents that create routines? Now, if you ask that to your chat GPT, you're going to get a very convoluted answer that that sounds very complicated. In essence, what it is, is that my recommendation, because I'm going to give, I don't want to leave you guys on a cliffhanger, is my recommendation is always, you need some form of database that lives atop all of your agents. And when I say database, it sounds like a complicated thing. But in essence, what I'm talking about is a spreadsheet initially. I personally use Postgres and Convex for all you devs, but the reality is for most people, a Google Sheet or a Excel Online is fine. It's actually better to start with a more simple thing than a more established. Security is not your problem when you're developing something. Security becomes an issue when you want to sell that thing or give it to other people. Before then, it's better to get something that works than to get something that's overly secure because your agents are going to actively work against you if your goal is security. They are hard asses when it comes to security and they're very annoying to work with. So security comes afterwards. But yeah, in order to have these cascade effectively, in my opinion, in terms of an audit trail, is once this is done, it's going to go and add a ledger entry. So once it's done doing its work, it's going to go ahead and add a ledger entry up here.

52:31And then you're going to be able to use that as the trigger for the next one. So when there is a ledger entry based on this active that's what's going to trigger the next agent if you want like a multi -step workflow okay so building trying to build multi -agent systems into cursor or whatnot is too complicated it's not at my level it's too complicated yes and no like cursor so all of this here, all of this here lives within something like cursor or codex or whatever. So you see all of the like the process with regards to managing the SDKs. This is where we come back to like our easy path, right?

53:24All of this exists within cursor. The only thing is you need to add on something of a audit trail for you. So you need to add this on. But this is literally just a connector. I think you can absolutely do this without any developer knowledge within the existing tools. All right. I'm going to I know I love it that you guys are so enthralled with this because this is exactly what I wanted. Sorry, we're just getting warmed up now. I know, right? I love how you're having this, Simon. It's almost like we're getting leadership skills for AI, you know, like the art of delegation, the art of management, the art of, like, strategy. And, you know, I personally get excited over that.

54:16I love it. Oh, I love it. You can do so much with AI. It's amazing. It is. It is. It is. So any other, Abhinav, you've been quiet. Yeah, because I had background music from my daughter going in, so didn't want to risk it. It's appreciated. Wonderful. Thank you. Thank you. This has been, I am already going back in and probably fixing up a couple of agents that I probably gave wrong instructions to. Not wrong, but not as insightful as yours. So thank you. I already started fixing one. Awesome. I'm super happy. Okay.

55:04Simon, if I could jump in. Thanks for this. It's a great session. I wanted to, you said that you do most of your work through, through codex and I totally get that. But when you're decomposing these sort of more complex tasks into, or the complex projects into their component tasks that you've got there, you kind of did that naturally. Do you use the chat to sort of brainstorm with you, how you break those down and then take that information and bring it to codex or do you actually decompose the sort of more complex task complex project into its component tasks in codex itself great question i love that question um i do not go back to chat i try to keep all relevant information on my brainstorming and thinking process in the same chat the reason I do this personally is because every single chat I do on my computer is logged to a database and I have meta agents that are actually going through every single session and understanding what was done, how it was done, why it was done, if I was talking to someone, what engagements I took, etc.

56:13If I didn't have that, I would still do it within the same environment because a it's more efficient and I think that there's a big element of not meaning not wanting to break your workflow. when there is a situation where the workflow is getting too complex. This is actually a very good idea for you to make a note, create a new chat within your like organizational repo and start having a strategic discussion of what it is I want to accomplish here. Oftentimes, this is an indicator that you have not yet sufficiently defined what is my business objective. And so you need to actually take a step back. So to rearticulate the thought, because I was kind of just like stream of consciousness there, but when I'm doing a workflow and I have multiple steps of it, if it's not coming naturally to me, it means I need to think about what does this actually mean for my business? if I'm doing social media posts, if I'm doing some element that requires multiple steps, I need to understand the who, the what, the why, the when, the how. And that will actually much more clearly articulate how much, what are the steps involved? What is the SOP involved? Knowing that, again, time is going to be the best judge.

57:29Once you put this in motion, after a week of it running, if it's not producing the results you expect, that is the eval. That That is the thing that's letting you know you've not sufficiently defined the process. Cool. Thank you. Maybe a separate topic for future AI and context. I've been asking the question lately is sort of if you think about like a carpenter and they build the wrong workshop for it. They build a workbench in order to be able to be really good as a carpenter. I'm wondering about what sort of workbench that you have that you're using as you build your AI projects and I see that you've got a whole bunch of stuff you've got auditing tools you've got this log thing you've got a meta meta prompts and that sort of stuff and I think that that's fascinating that would be a nice conversation to have in the future cool I'll do I'll do a session on my dev environment for sure.

58:27I'm happy to share. Thanks, man. Wonderful. Anyways, we're over time already. If there's any other questions, go ahead. I'll stay as long as people need, but feel free to leave if you have other obligations. Simon, I just wanted to thank you because it was always really fun to see you, see your thought process at work. And although it's a bit of a deviation, I would be interested if you once covered a topic of PRs more. That's because I have a hard time. And I can see the added value in Teams. And I haven't been able to see the real added value in doing my own work.

59:18unless you let agents pick up the PRs, then I can see where it becomes handy again. But I would like you to, well, spend a session on that type of stuff. Awesome. Yeah, I'll do that. I'll answer like very briefly still that because I think it's really great. Bye -bye Neilda, it was a pleasure. At the end of every working session, I commit. if oftentimes I'll commit during as well but at for sure at the end of every working session so every thread every day try to commit something because at least there is like there is a visual trail of I did this work on this day and agents can go back and refer to it cool Abdou pleasure as well wonderful so anyways yeah I do my work with AI here awesome thanks i'm gonna send stuff all of your way i'm gonna resume this i'm gonna recap this i'll send documents your way so you can browse it and no no no you are not gonna do that my agents will yeah that's that's a good ending i think take care Simon perfect thank you fritz yeah Thank you very much, everyone. It was his pleasure. Thank you.

1:00:45I just wanted to quickly ask how the second weeks will also just be posted also? Or is that like a... Yeah, for sure. I should have done that where everybody was here. But if you go to Luma, if you go right here to Luma, one more back, back, right here. Back, Luma, calendars, AI in context, calendar page. There you go. So if I go over here and I share this, you have all of the sessions I've planned already. So next week is very much about setting up your workspace. So it's very much about how do I delegate? How do I actually build these things out for myself? The following week is building lead lists, which is a topic that's come up by a ton of conversations I had and how to automate that entire process. And then proposal generator, multi -agent setups in the cloud comes later. And then I've got just like this general purpose session that's probably going to evolve. I'm open to changing things. I've already got these set up for myself. So if you do want to subscribe to the sessions that are interesting to you, feel free to do so. let me just drop that into the chat right now there you go i just wanted to check because i i just saw it through linkedin uh so i didn't know if there was like a schedule but yeah yeah thank you i i'm i'm doing these weekly i'm i'm over committing but hey that that i think it's a great way for me to i want to give away everything so eric no i appreciate it thank you I appreciate it because like I'm always trying to learn you know I'm doing my masters so like any knowledge on this is helpful because it changes so fast yeah agreed thank you again no problem perfect pleasure everyone see you guys next week I hope see you happy now