How I manage 250+ AI agents for a hedge fund

Hey guys, today I wanted to go over how I'm currently managing about 250 agents generally. That is the number of kind of active sessions that I'll be running at one time for an agentic hedge fund. Just to answer from the very start, I have three subscriptions on GPT Pro. So that's a $200 per month plan. I have one subscription on Claude Max and I have one subscription on Kimi. I'm gonna walk through that and then a load of other stuff that I've been kind of walking that I've been building and working on tips and improvements in particular since like the last video that I recorded for advanced workflows with Aagentic Engineering.

This isn't just specific though to engineering like in a load of the stuff that I'm doing. It's kind of finance related related on kind of back testing or just or jobs within an organization. So it's not just related to building things. It's kind of about building teams building agentic teams that can replace employees. So I have no employees, no external consultants, no other agencies. Everything done within this hedge fund is entirely done by myself and just through orchestrating large teams of agents basically. So let's get straight into it. So I very much look at look at agents as if you're you're hiring a person.

So, anyone who's built out teams, I think, is going to find this so much easier because my process has always been if I'm like hiring in a a business, it's because I have a need that I need someone to solve for me. I think it is a simple process of identifying problems that you need to be solved. Before, you would look to hire someone to fill that position, but now you literally don't need to do that at all. I'm not saying it never makes sense to hire people, but generally the hiring process of using a person to fill a position goes, you have a problem that you need to solve, a position that needs to be filled, someone to do a job.

You advertise the position, you interview candidates, it can take months. I worked in recruitment. It was my first job when I was 18 till 21. So I know that process very well and how bad it can be. And then you need to onboard that person. Whether it's a developer, they need to get used to the code base. there's always a process of onboarding. They might not like the job and they drop out of the position after a couple of weeks, after a couple of months. So, it's just a broken process basically, but it's the best one that we've had up until now.

So, you're looking at several months generally before you start getting full capacity out of someone. Whereas with agent hiring, which is exactly how I look at it, I look at it as hiring a team of agents. And that's actually what my skill is called. It's called agent recruitment. You have a problem to solve. You write a skill. I do a brainstorming process and like I said it's called agent recruitment so the agent understands from me what is the exact problem or workflow that I want and how do we break it down into different positions.

Then I will run that skill normally as an automation but depending on the skill I will run it a few times and just kind of monitor the outputs of the particular agent of that particular skill that's being ran. I will kind of iterate and improve. I quite like to use Kimi to do this at the moment. By the way, I've got like a Telegram group. I'll put the link down below. It's like completely free, but it's just for people that are building stuff with AI. I was suggested to use Kimi K2.6 for writing prompts and it is excellent, but it's also really really good at this process.

So this whole process I have it all managed by Kimi K3 who will it will interview me. I'll I explain the problem that I want to be solved and we will go through it and it will design a a skill that will then call out different sub-skills as I'll go on to in a minute and then generally I use Codex for it. It will be set up as a Codex automation which is just a cron job. It just runs the skill on a schedule. That's literally all it does. So treat skills like SOPs.

An SOP is like a standard operating procedure. You can kind of Google it to get an example. It's just essentially an instruction manual that you could give to a person or to an agent on exactly what you want it to do. So my kind of infrastructure setup that I use for managing this entire setup is I have my personal MacBook Pro here which I basically use for active work. If I'm actively building features or things that need my input, this is my kind of personal workstation. So at any one time, I would maybe manage 10 to 20 different chats.

And even if you're getting towards 20, to me, it starts to feel like I'm getting pulled in a lot of different directions. So I prefer not to do that. But basically, this machine here is for active work. Then I have a Mac Mini that sits behind me just over there. And that is my workhorse machine. at least that's I call it. And that runs about 150 automations, but they're basically workflows that I've come up with that are valuable to the things that I'm building and they just work on repeat. Some of them they might run once per week, others can run multiple times per day, some of them run hourly depending on what the job is.

Also, in some cases, I will run automations on my local machine here. When I say automations, it's just a skill or a prompt that's being run on repeat on a schedule. That's literally all all it is. But then if I'm running locally on this machine, I would look to move them over to the Mac Mini or I'll have an agent do it. It can just SSH in and migrate the automation to keep a kind of clean workspace here on my personal machine. I find it's much easier to think when you have a clear workspace.

And then I've got a M2 Max, which at the moment is in Georgia. It's just an old laptop I forgot about but have ended up bringing back into this infrastructure setup and that also runs it kind of depends but 100 agents that and that is my research it's my R&D department so to like not to go into too much detail about my project because I want this to be more like globally applicable but I've basically built out like a a back testing API it's ran millions of different back tests and those agents work on M2 max in an isolated environment.

They run the back tests locally. They analyze the results of all of the different back tests I'm doing. They can come up with new filters and kind of configurations for the different back testing algorithms that I'm running and they just constantly iterate a and improve. So that's like a whole other workflow that I mean it would be a whole video in itself but I'm not going to make a video on it because it's not something that I want to share publicly but it's basically like a research department but again it doesn't have to be applied for like a hedge fund like you could apply it for marketing for just kind of thing that requires testing.

I just have agents iterating towards a goal. But I will also say it's quite easy to do for finance related stuff because the results are quantitative like the agent can see if it's made an improvement using this strategy or this other one based on draw down final returns of the portfolio and a number of other parameters and a scoring system that I built out. So agents really work very well. it's very clear what they need to optimize towards particularly using the goal function in Codex. Anyway, for in terms of writing skills like I really like sentry has a skill writer skill that I'll I'll put in for kind of designing these workflows and brainstorming the superpowers set of skills also has its own skill writer in there which I do use as well.

When I use skills, I like to use the main agent as an orchestrator. And so he will then go and spin up different sub-agents. So sub agent one will then report back and then sub agent two and so on. So I like my process generally is to keep the main agent to have as clear of a context window as possible and it just orchestrates sub agents as a manager. So the agents do their best thinking when they have the most free space their context window. They need some context of like code base and what it is that they're working on.

But as soon as you start to compact, it is getting better. I will admit that. But as soon as you start to compact, the outputs generally get much worse. So I try to retain the orchestrator to be as intelligent as possible and then he uses sub agents to do all of the actual work and that is a general flow that I follow through everything from development through to research through to designing of new algorithms that we run through the API. Visually I kind of look at this as like a Zapier or like make.com workflow or like n for example.

you would think of all of the different nodes and each node could be a different so like the workflow would be the skill itself and then each step in like a Zapier or a make.com automation would be done by a sub agent right way of looking at it in terms of setting up a agentic department I give them a shared memory this doesn't need to be over complicated generally I just use a JSON file which acts as a ledger so a ledger being a history of the work so if an autom information runs every hour.

Let's say the first thing it will do is look at the ledger, look at what the last agent did and what the agents over the previous week did just to not overlap on their work basically or in some cases it depends what you're doing but it could even continue from there. But yeah, it just has a previous ledger as a scope of work and most of what the agents are doing are running skills on repeat. So examples that are going to be very common and more like generally applicable the most simple one would be find a bug in the code go and look through the whole codebase find an issue fix the issue open a PR and leave it in draft.

So that is my general workflow is I have them open draft pull requests and then I have a separate process which is actually an automation as well and could be in this list which is for reviewing and scoring PRs and it scores them based off of risk based off of is this just a monitoring change and because there's certain things I will allow them to merge autonomously and then there's other things that need explicit sign off and approval from me monitor issues in sentry or better stack. Sentry and better stack are just monitoring platform.

So if there's a bug or an issue in the code, it will get sent over to these so that they collect it and they've got a dashboard and MCP server and so on. So the agent can just go and check are there any new issues there. Create a fix, open a draft PR. Very useful. Fixes a ton of stuff for me autonomously. So again, these are just little things that are needed within the organization that would otherwise take me time to do. audit all Codex or if you're using claude you can audit all of your claude or any agent chats let's say over the past seven days and it can look for kind of common blockers that agents are having.

So it can normally be something like environment setup but it can be all types of stuff conflicting things in like docs compared to what it's actually seeing in the codebase conflicting instructions anything. I just like to run through and kind of monitor are there things that the agents are kind of wasting tokens on that could instead be added into documentation or a skill in order to kind of stop this happening. So I was also getting all types of things previously in our testing suite actually agents were just kind of running into problems and because this runs autonomously I'm not monitoring these chats I wasn't aware of it was on another machine and this was able to identify it and fix that problem.

So it becomes a self-improving system. This is my personal favorite one given the scope of the work that I'm doing that it's making investments, it's using real money, everything works completely autonomously. I mean I look over what's happening but the agents are able to do things autonomously. So the most important thing is to make sure that the code is working as expected and the other most important thing is okay so if you deploy the code it may be working at that point in time but enforcing that there is no drift that things do not silently change a week later a month later a couple of months later basically the agents they could accidentally make a change that could change the outputs in how my trading system functions and I need to make sure that doesn't happen.

So in order to do that, it does regression test with something called golden inputs and outputs. So if I give it a fixed set of data, let's say a fixed set of prices over these dates and it runs the trading system, those outputs should not change. It should be static. So it's a simple test of if I input this data, the output should not change. So it gets frozen. So I have thousands of these now testing every single different configuration and option within our you just wouldn't think of like the number of different like possibilities or different areas within a trading system.

There's just so many things that if they move even like a tiny percent it changes the outcome of everything. So everything basically gets frozen. So this automation would look through it would go and look for a critical code path generally related to to to trading itself and it will then freeze it. It will generate kind of like dummy data like demo data and it will get the outputs and it will freeze the test. So if ever the output that the code gives us changes, it won't allow it to be deployed into production, unless it explicitly gets a human approval from me which is done from a second admin GitHub account which is only available on my phone.

So the agents don't have access to it to be able to bypass it. It needs explicit sign off from me from an external device. Yeah. Yeah. I mean you can even read over these. One other concept I came up with as well is to come up with behavior validation contracts and monitoring agents. So again, it's now so easy and so cheap to write code speaking at least to what it was before. So it's one thing to like you can have these agents pushing code into production all day long and I I mean I'll make a whole another video on this, but I actually think it's a massive problem of it's so easy to do.

There's no friction now that you can just deploy, you can just push a load of crap that you don't really need into production. One thing that really it's so easy to write code, but then in terms of actually validating that that code, you know, it may have been working fine locally, but validating that that code is working in production as you're expecting it to becomes another problem, especially on systems like what I'm building where there's just hundreds of different moving parts. So there's a couple of things that I've come up with to solve this problem.

One of them is behavior validation contracts. So all of the critical features functionality, it could be cron jobs or workflows that I'm running within render, they get something like a behavior. It's a behavior validation contract. I'll show an example of it here. So this is a visual example which is something that I've been working on but it's actually just a YAML file. So it's just plain text and so it would be something like step one it's expected that the cron starts and makes sure that everything is working. It can go and fetch something from the database.

It's basically laying out the exact steps. And I've been trying to turn this into something like a Zapier or like make.com style thing, which is what I was replicating here, but currently it just like it will look something more like this notion document on the right hand side. It's basically just writing out the steps that are expected to be performed in the code. And then you give that to an agent who would then go and validate to make sure that everything is working as expected. And in my behavior validation contracts, I would ask them to validate for a week.

In some cases, it could be a month. And you would give a cadence. So, how often do you want this to be checked? Do you want it to check hourly, to check once per day, once per week? And again, it depends on what it is that you're running. But an example of this would be like I currently ingest comments and information from about 50 different data sources from Reddit, from Tik Tok, from Instagram, from different news articles, from different finance publications, various different tools. which is what the investing theses are built on top of.

So a behavior validation contract could for example say on Tik Tok make sure that we are correctly absorbing all of the comments from all of the sources that we want to take them from and that they're correctly being saved in the database and then that our AI sentiment analysis pipeline is running on all of those comments to check whether they're positive, negative or neutral. That would be an example of a behavior validation contract similar to site reliability just monitoring that all of the different systems within the organization as a whole are working.

So each of those have a behavior validation contract and then a monitoring agent who is assigned to make sure that that contract is being fulfilled. That's just an example of how it works. So previously maybe you would have needed people to do it. You could have potentially done some of it with code, but I think not in the same way, especially now that you have the computer use ability within Codex, which I'm just a huge fan of. For anyone who hasn't used it, you should go and play around with it.

It's an unbelievable piece of technology and I use it for all types of stuff. So basically it allows the agent to use your machine to use your mouse and keyboard which means you can get a lot more advanced and intricate with the type of validation and monitoring that you're able to do here. Anyway, another example of an automation would be look for code that can be simplified. Again, I'm really big on on this phrase down here. So as complex as necessary, as simple as possible. This is something that I've just been obsessed over for the past couple of months.

And this is an example of how to autonomously apply that principle of always looking for overengineered code paths, overly complex code paths, and to simplify them and open a PR. Look for superseded code, so code that can be deleted, that's no longer relevant, that can be removed. And then a few other ones that I find useful are like scan Codex chats for user preferences. I've been using Codex now for the past November last year. So, I've got like such a wealth of data there that just sits there in chats statically, which does not make any sense at all.

The thing that I like to use for this is is Kimi has a swarm feature and it can kick off up to 128 agents and it can go through and read all of your old Codex chats and it will basically mine for user preferences for it can look for design decisions. decisions that you've made and they can then be put into a document. And the reason I like to do that is just because it stops you needing to repeat yourself so many times. One of the biggest headaches that I used to have before was that because the chats are kind of isolated.

You tell one agent and you give it all the information, but then it's only that agent that has that context. And this kind of solves that problem by Alexa like a swarm to go over all Codex chats and mine design product decisions and put into a document. Anyway, I'm not going to do it cuz I don't want to expose all the stuff that I'm doing here. But yeah, you would just have it make a document and it can go and commit that into the codebase and then all of the other agents have access to it as well.

And again, you can just schedule that to run weekly or else. And so yeah, that's been like one of the biggest things for me. Again, what I'm talking about here is and what sorry, what I was going to saying about it's one of was one of the biggest headaches for me. What I'm really focused on now is these agents can basically work 24/7 and the limiting factor becomes you as a person. So, how many decisions you can make per day as in pass that back to the agents. What are the quality of your decisions is a massive I think the biggest thing now because you now essentially have infinite leverage through AI.

So, it literally all comes down to the quality of your decisions and going in the right direction, which I guess is somewhat subjective, but sure that you're making high quality decisions. Um, but basically, I'm fully focused on trying to have the least amount of stress as possible, not needing to repeat myself because repeating myself to these agents, it's a waste of my time, but it also causes decision fatigue. So the speed at which you can move is now basically down to how many decisions can you make a day and communicate that back to the agents to allow them to actually do the work.

And so I'm so focused on building out systems and kind of workflows where I don't need to repeat myself and they can kind of look for the answers and if they have very high confidence of it, they're able to kind of proceed autonomously without needing me being the blocker because the blocker does become the human in the loop. you are still required but essentially you should be looking at your job. I just look at my job as like a product owner basically. I'm the owner of the product. I choose the decisions in terms of what direction are we going in but other than that I literally don't get my hands dirty at all anymore.

I haven't written a line of code in I think since I downloaded Codex. I do review PRs. I do look at PRs but that's much it but I try to do it much as possible. So what I've listed above are mostly like development workflows. Some of them are kind of more for more value out of the agents. I have a whole R&D department research and development where they research new back tests or new structures for algorithms. I've built out the infrastructure of this entire back testing system and then the agents just have two or three skills.

One of them is on how to run a back test, how to send the API request in our system, how to come up with new ideas for back tests, how to analyze the results, which is kind of just my thinking put down in words of the types of things that I think that they should look for. So given my background in growth hacking, we very much apply a growth hacking or creative thinking approach on how to think outside the box to find correlations on investment opportunities. trying to get an agent to think outside the box.

It's an interesting challenge to say the least. Not a problem that I would say that I've fully solved, but we're getting there slowly. Anyway, one other piece of advice I would give to keep in mind, I've touched on it and I'm going to make a full video on this because I'm just this is so smart. As complex as necessary, as simple as possible. Do not add unnecessary complexity into your codebase, into your systems, into your business, into your life, into your fitness. It just applies to anything. It's so simple. It's so smart.

Keep everything as complex as it needs to be to solve the problem, but do not add anything unnecessarily. Keep this in mind when you're building things. Your agents try to add constantly. They try to add more complexity than is needed. And your job or you need to have systems in place to do this needs to be to simplify to refine to its most to the solution's most simple version. To not add more complexity. More complexity means more code, more tests, more context bloated because agents need to search around your codebase for different things.

more failure points, more things that can go wrong, more time in order to do things. It's just you like this is the main thing that I spend my time fighting against is having a minimal code base because the things that we're doing are quite complex anyway. It just ends up being like spaghetti code or like a spaghetti project if it gets out of hand. So, I spend most of my time really on trying to keep things in shape here. Anyway, that is like the first part on kind of agent automations from let's say like a high level of what I'm doing there and how I'm getting value there.

And again, this is like real world value. Like my outputs are I would say like 5 to 10x what I was in fact it has to be way more than that. It's very hard to quantify. But all of this stuff that I'm doing now would not have been possible for me a year ago even to the extent and at the scale that I'm doing now with all of the things that we're doing with the machine learning with distributed hyperparameter optimization which I didn't even know existed and now there's like a whole team just dedicated to that on training models.

This opportunity would not have been available to me a year ago. And again, the thing that I find so interesting about AI, we all have access to these tools. I'm using the same tools that you are. Like GPT 5.5 High, GPT 5.6 Soul, Kimi K3, like I'm we're all using the same stuff and it's completely down to you. How big of a problem are you going to assign these models to? Because they will eat up pretty much anything that you throw at them. At least from what I found, they get confused.

But again, that gen generally tends to be down to the person managing them on how good your prompting is or on how clear you are on what you want to actually be built, which I'll go on to in a minute as well. There's a whole other kind of workflow there as well. Anyway, the next thing that I'm going to introduce you to, which is my favorite or one of my favorite tools, is Codex app server. Basically, it exposes an API. This is officially built by OpenAI. It's built into Codex. So you would just run the command Codex app server and it's going to start running that server and it gives you endpoints to be able to actions within Codex apps.

So you can send a message, you can rename a chat, you can archive a chat, you can create a new chat. And so here we have Kimi and I could say please list open chats on Codex server. Kimi is going to be a little bit slow but it will do it. Like it's a bit slow to set up but once it's done it's pretty fast to do. Basically, this is how you could have Kimi or in fact, we can even do Claude. I would probably use it to like triage chats that are blocked and need my attention or chats that have been superseded and can be archived.

So, I I often just kind of let that run in the background. I've already ran this this morning, so I think I'm going to have less than 100 chats open on this machine, but I I've had times before it's gone up to like 700 or 800. And on that Mac Mini is even more. It just becomes impossible to manage yourself manually. So, what this does is it's listing out the Codex here and it gives Claude access to all of your Codex chats. So, I could have Claude concurrently managing tons of different agents.

So this is a great tool. Again, it's built officially by OpenAI to be used exactly for this type of purpose to be able to have like orchestration agents controlling an army of codec sub-agents. The issue used to be that Codex had a 5hour coding limit and when you would do this you would hit that limit really quickly. If you were to spawn 20 or 30 agents you would be running into your 5h hour Codex limit. Whereas now they've got rid of the 5-hour limit, which for me has been a massive unlock.

So now the only thing that applies is the weekly limit. I don't know if that's for all plans, by the way, but it is the case for the pro plan. I'm just going to click away from this because it's going to show a load of things I don't want to show. That will give it access to all of my Codex chats and I can basically manage hundreds of agents. I mean, there's no limit to the scale that you can run at by just communicating with one Claude chat or one Kimi chat.

I like to use it for identifying blocked work or work that needs my attention. On top of that, you can also use it with Hermes agent. I really like Hermes agent mostly because I can use it from Telegram and Telegram just feels so native to me and I can do voice notes with it. So, my Hermes setup is using Kimi K3 with Hermes. I'm not sure if it's available on there on the API now, but when Kim K3 first came out, it wasn't. And so I use a tool that's called Vibe Proxy.

It's open source that you can use and it allows you to use orth basically anything. For example, Gemini won't allow you to use your subscription with Hermes. So you can use like Gemini Flash by using this tool. You can use your Kimi subscription authentication if you route it through via proxy. So yeah, I basically like to use Hermes as a chief of staff. As I mentioned, I've I've got three different machines running that that manage all of these agents. They're all completely isolated, but each of them has their own Hermes installation on and it uses Codex app server.

So I basically spend every morning walking up and down the beach for 2 hours just sending a load of voice notes to these chief of staff agents. Chief of staff it's a a kind of traditional term for someone that would ma manage all of your staff. So I use Hermes as the chief of staff, but in this case it's like chief of agents, let's call it. And I I just use voice notes just to kind of communicate when I'm out out and about. And I I think it is a great setup.

It's the most value that I've ever got out of Hermes is is since Kimi K3 came out to be completely honest except for before when you could use like Opus 4.6 on it. Okay, the next thing that I'm going to go over is looping over plans and iteration iterating on on your plans I found to be one of the biggest unlocks in terms of quality of outputs. I use Ora superpowers. He's a really interesting guy actually, but I've completely forgotten his name, but you can get up him. one of the most starred repos on GitHub, but he did a podcast and it's just really interesting to listen to his thinking and thought process in terms of how he thinks about making skills.

So, in a previous video, I had my own workflow that was an implementation workflow, but I've now moved over to using this superpowers. Probably about a week after I put out that video, in fact, I discovered this and moved over. So, they have a skill called brainstorming where it basically interviews you asking what would you like to be built? It's really good. But funnily enough, in the podcast from the creator of this thing, he said if he wanted to make some improvements, it would be to do iterations on the planning.

But he said he hasn't done it because it would take much longer and he thinks he's now kind of like building this for the masses rather than for his own preferences. So I listened to that and I thought, you know what, that is a great idea. And so I've built it out myself. again just using his thing, but I've just built out a new kind of loop that loops over over the plans. So, what I do is I launch two adversarial agents to audit the plan and to look for gaps and weaknesses and things that the original plan missed and things that should be clarified.

It basically always finds something. Now, most of the value is generally in the first two to three iterations. So, it it's going to depend on what you're building. Again, I'm building like finance products. I mean, not products like they're they're only for me, but it's finance related. So, like attention to detail is just critical. If I'm working on the critical area of the codebase, I'll do up to like 10 iterations to make sure that everything is completely I prefer to be over specified in the plan so that it builds what I want and doesn't puck anything up rather than it builds something and then I have to spend ages using agents to fix the PR.

So this will go ahead and it will find a load of stuff. It will ask me questions. I will reclarify them and then I will launch again two more agents, launch two more agents. So every time there's an update and we update the plan, agree on plan, update the plan, then we launch two more agents and then it will find things. It asks me what do I want to do? We agree on it. It then updates the plan and then we launch two more agents. So that is the iteration cycle.

And again, the fact that I'm able to do 10 iteration cycles and it's able to keep on finding new things that should be clarified, it's so so valuable. So again, 10 can be like super excessive, but two to three, it's going to take you much longer. By the way, I I've in some cases I've spent about 2 or 3 hours on a plan. And by the way, this is after the plan has been returned. So the first agent has given the plan saying this is ready to be implemented. are you happy to proceed?

And I then say, wait, let's launch two agents. And so the plan has already been made and the first agent believes that it's good to go. And in almost all cases, it turns out there's like a ton of improvements that can be made. So run this iteration loop over the plan. Also, just to clarify, each agent that audits the plan, it's a fresh agent every time. Every time I launch an agent, it's a completely fresh context and fresh pair of eyes. So, I don't continue from the old chats. It needs to be completely new agents that are reviewing the plan.

So, yeah, it's going to take considerably longer, but it's just best to do this process up front to clarify what you really want so that the agent can actually build what you need rather than have it guess at things or you end up with just a load of crap that you don't need. So I launched two audit agents on each run. Like again, you probably could do one. My mindset at the moment is just to throw tokens at problems. It's like tokens are relatively speaking, they are so cheap right now for the amount of intelligence and the amount of value that I get that it just makes sense to throw tokens at problems.

I personally GPT 5.6 Soul I for building stuff it's good, but I hate it. I like and I'm a fan like as everyone knows like everyone couldn't believe that I was saying how good Codex was five or five months ago when no one was using it because everyone was obsessed with Claude. So I've been like a long-term OpenAI user but I'm not a fan of talking to 5.6 Soul. I find it's not good at planning at the kind of design of things and just generally to speak to I I just don't really like it.

So I I generally speak to 5.5 high, but for auditing and they I find they think in very different ways and I find GPT 5.6 soul does miss a lot more despite it supposedly being a much smarter model. So one agent on 5.5 high and one on 5.6 soul and I often find that they will come back with different discoveries which is obviously good. You could also go ahead and swap this out for claude or for every else open code on some open browser model if you wanted. So, in terms of models, because I do get asked about this a lot and like my choices for different models, I really like Kimi K3 and now it's kind of really hyped out a lot, but when it wasn't being hyped up, it was just like a really pleasant surprise.

I had it rebuild a load of stuff in just like the internal dashboard that I have and really wellought out design from like a user experience point of view. really good front-end design, really well thought out. Considering the things that I'm building are really quite complicated in these trading platforms and how to display all of this data back to me on what is critical data, not overwhelming me with too much information, but really really nice model. So, anything front- end related, I like to use Kimi K3. I also like to use it for writing prompts and for all of my automations up here, I have Kimi K3 manage and build all of them.

And it's set up using a Hermes agent. And then I also like the swarm feature which I demonstrated the start of there and it's also my main agent on Hermes because it has a I would say a nice personality. GPT 5.6 soul high I use for building and implementation and some QA. GPT 5.5 high or extra high I use for planning architecture and for writing plans using brainstorming. Fable 5 I really like for ultra code. Also, if I'm looking to just kind of oneshot a side project or something that I'm just building for finite, I like to use alt and fable 5 because it will just go and spawn like 120 agents.

It's quite satisfying to watch. Opus 5 was released about a week ago now and I have used it a little bit and yeah, nice model to work with and I use for a little bit of code review. In terms of my actual setup at the moment for programming and actually running agents, it's predominantly using native Codex. I also use ghosty which is what you see on the left side here which I've completely moved away from warp. So native Codex this app here I like to use a lot and but I this Codex native app and their native mobile app is really good and I again I now spend a lot more time kind of out and about walking around but they've also released audio calls on the app here so you can now do calls with an orchestration agent I think it's Luna but it doesn't actually specify what it is but it's not a super smart model but it is fast at kind of orchestrating and it can basically manage all of your Codex chats on a call.

So I probably spent about 3 hours on a call to it yesterday going through planning. And the thing that I did find it's pretty good for is it simplifies advanced concepts. So if you find that you're getting confused when talking to the agent, if you're not sure what it's referring to or you need to make a plan, you can start a voice chat here. It just does a much better job at speaking more clearly. So, it's basically just dictating what I'm saying and it's just going to read back to me.

So, I won't do it now. But another amazing piece of technology is this computer use which people aren't really talking about. It's one of my favorite things that has been released over the past year and I use it for everything from ordering a coffee to it can control your phone if you use iPhone mirroring to debugging stuff on my machine to anything that you can think of. But now you can also do a voice call and it has computer use. So I can actually tell it to control my machine just by speaking to it, which it's just a new way of interfacing with a computer.

You know, we've always used like a mouse and keyboard or trackpad or whatever. Whereas now, simply with my voice, I'm just able to tell a robot basically sitting on my machine what I want it to do. And handsfree, it's able to control and do everything. Now, one step above that is you can also do it from mobile. So, from my phone, if you have the OpenAI mobile app and you click on remote, you can call your laptop. So, I can be out and about as long as you've got like a stable internet cuz if internet drops on either device, the call disconnects.

But you can be out and about walking calling your laptop telling agents what you want them to do and it's going to be launching like 10 to 20 like I mean you can change the configuration. I've got it set to like 24 so it can launch up to 24 sub-agents at a time. So again, this has only just come available in the past week and so I'm still like new to playing around with it, but yeah, it's just such a such an interesting time to be alive and such an interesting way to interface with with the machine.

Behavior validation contracts I've gone over already. And then one other improvement that I added as well was another adversarial agent review. When I say adversarial, it just means attacking. It basically means not to agree with what is written here. It attacks, it looks for weaknesses and challenges the thinking of the other agents. So now in every pull request that I have, there's a required field. It's just a checkbox, but it basically needs to be marked off and there need to be comments from the two agents confirming that they've signed off on that they're happy with the state of the PR.

I find that they find so much stuff, especially as like our code base is now getting larger and larger and there's more moving parts. An agent could change something that accidentally changes how another kind of module or part of the application is working without realizing. And this process is really good at picking that up so you can enforce it when you're using GitHub PRs. I've kind of just touched on this already, but yesterday I was literally walking around this apartment using a planning mode on voice mode for quite a complex system and I found it to be pretty pretty good and obviously gets your daily steps counts up.

Ponytail is a skill that basically makes your agents a lazy engineer and I like to use it quite a lot. Again, what I was talking about on as complex as necessary, as simple as possible. This is kind of a skill that follows in the in the general idea of that. It tries to avoid overengineering. It looks for areas of the code that can be reused. It tries to not add unnecessary code or or just yeah or complexity. Basically, caveman skill changes how your agent communicates with you. I like to use it sometimes as well.

One of the biggest problems and things that I'm still working on now is how to get the agents to focus as clearly, succinctly, and just concisely as possible. Um, being as clear as possible, giving real examples of things is something that I'm really focused on is how to make as many highquality decisions per day. And every time I have to go back to an agent asking him to reexplain it more simply or if I have to spend a load of time reading a long message, if I'm dealing with like hundreds of messages per day, all of that really kind of adds up.

I think some of the biggest productivity improvements to be seen are on it's something I came up with called decision packets. I won't get into it now because this is already a really long video, but it's kind of like a formatted structure that it should send to me and it gives me suggested options. So I can just put like A, B or C and it gives real examples of how this change would affect the behavior in production. For example, there's a tool that's recently come available called BuzzA, which is basically Slack for agents that you can sit in as well.

I've been playing around with it, which is why I don't want to open it cuz there's going to be a load of confidential stuff in there. But basically, I it's quite smart and I've wanted a replacement to Slack for a long time and nothing has really existed except for Microsoft Teams and I just can't stand anything by Microsoft. This looks even just as a replacement to Slack looks pretty good, but it's basically agent workspace. So, you can invite all of your agents and you can basically have Claude communicate with Codex and set up different rooms and so on in different workflows, but it's very early stages.

is by Jack Dorsey and starting to pick up some traction, starting to be spoken about a little bit. I've set it up. It's just something that I'm monitoring and keeping an eye on. I've been thinking about building a small lightweight SAS of something that I think is quite useful because I'm monitoring about 50 different platforms for all of their comments, which is how we make our investments within this main project. I could build a SAS that would be like an alerting tool for everything that you want to track. So the way that I use it is one for finding investments, but two, if I'm holding a position, I like to get alerts and updates on what the general consensus is of people at at the moment, let's say like on Reddit or on Tik Tok.

But technically, you could set up alerts for anything. So it would be like keyword based triggers, something like Google alerts, but instead it would be for social media, for monitoring comments and so on. and discussion on social media. So you could do this for something like if your only fans model gets mentioned on Reddit or something, I can send like an email or like a telegram message to that conversation of where that actually took place. And it could be something like leaked content. It could be a customer. Again, it can be an only fans model.

could be anything like it could be I don't know someone selling like running a peptide business and whenever someone says like where to buy peptides in X location in I don't know in Dubai or wherever I would basically monitor the whole of social media because it's what we're doing already for comments that and whenever we find a comment that matches that particular keyword you can basically get an alert about it again it could be for a brand it could be if someone's discussing your brand name in a particular particular community or like again it's Reddit, it's we monitor newsletters, I monitor Google, I monitor about 30 different finance applications, they wouldn't really be relevant.

I monitor Tik Tok right now with scraping hundreds of millions of posts and I think not quite a billion a lot of comments and running a lot of analysis on there to build out theses for investing. Anyway, I'm going to cut off the video now because we're approaching an hour, which is long even by my standards. There's quite a lot to go over here, but I hope it was useful and insightful on how you can start building out some of these different tools and how to get more value out of AI.

Any questions, as always, leave a comment. And as I mentioned, there's a free Telegram group that I've set up with about 500 people in for people that are building and coding with AI. So, if you're interested in that, feel free to join in the description as