Advanced AI coding workflows: agent teams

Hey guys, hope you're well. Hello. First off from the as you can see from the background from the Maldives. If you keep your eye on that background, there was actually some dolphins out in the the darker area this morning that I was I was watching. Anyway, today we're going to be going over some AI workflows. I'll be doing actually just I keep getting accused of using fake background.

So, just for the sake of it, I'm going to come out here on the netting to show that I do actually live this life despite what some people might think. We're going to be going over some AI workflows that I've I've been using for advanced use of Codex, Claude Code. if people are using cursor like this doesn't really doesn't really apply but it might kind of show you like some ways of like transitioning into the these more more advanced workflows.

So for full transparency I have a $200 subscription of GPT Pro GPT-5.2 high reasoning is my favorite model by a long way. I know that the industry is obsessed with with Claude Code and I think partly people a lot of YouTubers just talk about it cuz that's where the search traffic is and it's like a self-fuelling fire. I've been working with these tools every single day I think for over a year now.

GPT-5.2 it is just so much better at programming than than Opus. And I was an Opus fan. I also have still have a $100 subscription on the Claude Code Pro plan. I think it is one below the max, but you know, previously I was on the the max plan back when Opus 4.5 was was there. But Codex is the CLI tool of GPT, right? So again, just not to be confused because they they've named like three different products called Codex.

It's terrible branding by OpenAI. They have Codex the command line interface as in equivalent to Claude Code then they've got Codex cloud just to for transparency whenever I'm talking about Codex I am simply referring to the the CLI their alternative to Claude Code right and it just shows how badly they've done this I actually have to explain all of this at the start of the video the only model that I use for writing code is GPT-5.2 in high reasoning.

I have subscriptions to GPT to Claude and to Gemini and I've tried all of them and I do continuously get like a fear of missing out when people are always talking about Claude code and whenever I go back to them I run a lot of experiments of them in direct comparison. The only time that Claude code ever beats GPT-5.2 do is maybe on UI design accuracy, quality of code, following of instructions, just generally working with an agent like GPT-5.2 high reasoning wins by an absolute mile.

Codex as in the model I I'm not a fan of, but GPT-5.2 in my opinion is the perfect model by a long way. In comparison of the different models, I find that working with Opus, it's fast and because it does a lot of parallel tool calling and things like this, it makes it feel a lot faster than GPT does. But I find Opus is like working with an intern who's very enthusiastic, wants to solve the problem for you, is quite proactive, but kind of makes a mess whilst doing so, particularly on larger code bases.

And due to the limited context window, the performance drops quite drastically. I'm I I this is like an unbiased opinion. I've got nothing against Opus. I I used to love Opus 4.5 when it first came out and I thought it was unbelievable, but GPT genuinely has just overtaken them. So Opus feels like an intern. It will eventually solve the problem, but it will make a mess make a big mess whilst doing so. files all over the place, different test scripts or like not following instructions.

Exactly. Hallucinating more. GPT-5.2 is like a bit pessimistic I find and maybe that's why people don't like the personality aspect of it. But it will measure twice, find exactly what it needs to do and cut once. And it's very good at keeping context of the bigger picture in mind. So the changes it makes, it won't break other areas of your application. So I'm just to give some context as well, my current main focus is making a billion dollars with an AI only team of investors and I'm doing this through scraping Reddit.

I've like come out of retirement to basically focus on this. So I'm basically building like an AI only hedge fund or or at least attempting to and it's just quite an interesting project. And then the purpose of a billion dollars, it's just a way of kind of measuring things. It's not for anything else. It's just something to motivate me a little bit. So given what I like this is using real money, making real investments, GPT-5.2 again, like Claude just kind of makes a mess and I I don't want it being involved in anything related to real money.

Only thing that I use Opus for is every week I kind of do a comparison between GPT-5.2 and and Opus just to make sure I'm not missing out on anything. I do use Opus in OpenClaw actually by the way as like a personal assistant. But again talking specifically about programming, I would encourage all of you to just try out GPT-5.2 high reasoning through the Codex CLI. It also hallucinates a lot less as well.

Right. Claude Code verse Codex that stuff my current AI workflow and just about the environment like I work entirely from GitHub worktrees now you can ask you know GPT or Opus or whatever you want what are GitHub worktrees but that is the environment that I work from so that you can run multiple agents in parallel which is how I'm able to to build things out so quickly and the these are the only apps that I use at the moment by the way Safari I've gone back to from Zed and from Arc I've come full circle and ended up coming back to Safari and actually really liking it.

I use Warp as my terminal manager for managing well, it's got AI built in into the terminal. And then finally, Codex here. This is the actual app. So, for anyone who isn't that technical, this may actually be a really good tool for you. I prefer to run them myself in the terminal and have a little bit more control. But for anyone who isn't that technical, this is actually a really powerful app and they're very actively developing this cuz I just updated this this morning and it's now 12:30 in the afternoon and they've already got another update out.

But I'm not going to go through this because I'm not using it. If you kind of look at something like this and it's a bit intimidating, I would say that this Codex application is like a really good in between in terms of being a big step up from cursor and the capabilities there. You're using a direct openai product. So you also get they've currently got an offer on of like 2x usage limits if you use their products directly.

So you'll get a lot more value for money than you do out of cursor. And I just find generally that using the products directly are a lot more kind of powerful than using variations like through cursor or through the these other tools. So that is what I would encourage you to look at this. I have been playing around with it but again I prefer my main workflow. So just to give you like this is my current workflow by the way.

There's been a ton of iterations on this over the past kind of 5 to 6 months and it's subject to more iterations like I do constantly change and improve the workflow but again just from like a high level I'll go into this in detail in just a second but I have an idea to build something you know I want to build a dashboard I want to build whatever I'll go through a real world example in the actual terminal in a second I have a template for a product requirements document and requirements clarity so I'm very a big believer in like this spec-driven development.

So this is the only process that I'm actually involved in. This is where I'll speak to an AI agent. We go through 10 rounds of questions to basically really clarify what is it that I actually want to be built? How is it going to be used? What are my expected features? And it will kind of interactively ask me all this stuff. So it's got 100% clarity on what I want to be built. Because the biggest waste of time is if you don't do this and you know the AI misinterprets what you want.

You start going in the completely wrong direction. Your code's been changed. You then need to clean everything up. You miss stuff. You end up wasting a load of tokens on things. You've wasted a load of time. And like again, this is talking from experience because I've been over this workflow so many times. So building a really detailed and comprehensive spec of what you want to be built even like over giving on details of what you want to be built running through this workflow with this one agent who will then translate this into multiple separate different GitHub issues.

By the way, I work entirely from from GitHub issues. So everything gets written up as GitHub issues over the command line interface and then I I mean I actually assign them to an implementation agent who then launches sub agent. So it will launch a dev agent. Once dev agent says he's done and everything's satisfied, it gets launched to a separate QA agent. The reason that I like to do this is because they all have different context windows and I actually get them to attack each other's work.

If the dev agent has done the implementation and he says it's done and then you ask him to run like a QA workflow against it, he's biased. He's going to say that it's all done. In actual fact, there's a high chance that he isn't. So dev agent finishes his work. The QA agent is then launched. Again, a separate agent, completely fresh context window, no past history of what's going on, and he's given a very long, first off, he checks the GitHub issue for the requirements, the acceptance criteria, everything that we're expecting.

He can read the epic that was created and he then goes through and does end-to-end tests. Does code checks as in making sure the code is running very efficiently that it's not kind of bloated or overengineered. Verifies that all tests are passing. Verifies that there's no unnecessary changes. But the most important thing is like the end-to-end testing because often I find that they try not to actually do end-to-end testing in like a real world environment.

But I'm very specific about if we make a change on Air Table or we make a script, you need to run it and you need to prove to me that it's actually working and you need to actually show that proof on like a GitHub issue or something like that. Often it will fail and so then the QA agent kind of gives a report of feedback and that automatically gets passed back to the dev agent for him to then go and make those changes. fix passed back.

New QA agent is spun up. Runs all the same checks again. Did was everything fixed? Is he happy with it? It gets passed to the merge agent whose job is to do we have AI code reviews. So I'm using cubic.dev and code rabbit which are like AI reviews. I'm also using another it's a linting tool. I think it was called Ruff. The merge agent also updates the relevant doc. So if the changes that we've made, if that needs to be reflected in our documentation as well, it's also the responsibility of the merge agent to do that before he merges that branch into the main branch on GitHub.

This process is all done automatically. I'm kind of listing off the workflow for you, but it's all done with one command that I'll show you in a second. So the only process that I'm actually involved in is up here in the product requirements document which is why it's so important that this is like really detailed and really comprehensive because after that it's getting implemented the QA agent verifies that we've built what we want and then it's getting put in into the main branch to be ran.

So I know some people are going to say like I think anyone from a more traditional programming background is going to be very anti this but again I'm just kind of trying it out and so far it's been working and I like it. So yeah, update docs, close the issue, and everything's been merged. So that is my current development workflow. So just to show you how this actually works as a real example, I'll make a new warp window.

We run the shortcut PRD. And this is going to run my workflow. So you can see the prompt here. You are a PRD epic agent as part of an AI only engineering team. I'm not going to read all this out to you, but I can put it in the description if it's useful to people. You can take what you like and remove what you don't. So the main point here is its job is to get us to above 90 clarity which is like a kind of score that I've come up with.

And the way that it does that is it does rounds of questioning to gather information from me on what I want to be built, how I want it to be built, expected behavior, how it's going to be used, exactly what I want it to do, and it's then going to translate that into requirements and issues which as we showed as part of the previous workflow. It then gets passed on to these dev agents and they handle the implementation.

So the way it also does this because that I like it to do is also I tell it to talk to me in extremely plain and clear English. I have this written all over my prompting because I don't like it when it is kind of ambiguous and isn't clear in its writing. I want it to be incredibly clear. Right now it's reading over our engineering docs. You can see it's doing all of that down here.

And it's basically going to work with me in this workflow. So it asks me a question. It makes three suggestions on approaches of doing this and it gives its recommended option. Of course, I can also just reply to it saying no, I want it done in a different way if it's misinterpreted something. But this makes the workflow so much faster as you'll see in a minute when it makes its suggestions.

I can just write 1 A, 2 B, 3 C, 4. No, don't do that. Do this. Right? And it makes it so much quicker for me to be able to kind of work with this agent. And again, this is what I mean about GPT as well. It's going to spend a long time reading through all the docs and understanding everything, which makes everything so much better. So, you can see an example here. Right now, the clarity is very low because it doesn't even know what we're doing.

And it's asking what are we going to be doing. So, I can say I want to build some charts in our dashboard to give us a broad overview of the current market fundamentals. As I mentioned already, I'm building out like an investing tool. So, this is quite it's purposefully very broad. So, like there's no specifics of like what I actually want it to do. So, as part of this workflow, it's going to have to ask me or make suggestions on specifically what type of charts do I want?

What do I want to be tracking? If I don't know, it can make suggestions for me on what it thinks is a good option. So the whole point of this is if I'm not specific in what I want, this is kind of like a brainstorming exercise where we actually go through everything together, really hash out the details very very specifically of what is it I want, what am I, what is the expected behavior, and then it's going to translate that into a load of issues.

I'll show you how it looks in a second, but again, like it it will translate that into maybe like three or five issues that I can then assign to different dev agents. They then get added in here. And then the next stage would be let's say that this issue has been planned. So this is issue 1237. I can come in here and I'll just type impl 1237 impl is just short for implement. These are just different shortcuts that I've basically made using .zshrc.

Here you can see my prompt. You are the implementation orchestrator on GitHub issue 1237. So this is automatically going to set up a worktree which you can see it's done. It's made branch 1237 up here and it is going to say you must use this skill GitHub orchestrator which is basically just a set of instructions which I can actually open for you here where it says you are the orchestrator.

Your job is to basically just to manage the workflow that I specified for you up here. So first it will launch the dev agent which you can see launch github issue dev agent you don't do any work yourself um understand the issue everything like that launch the dev sub-agent wait for that wait for the dev sub-agent to finish then launch the QA sub-agent he runs the checks then you pass the feedback back to the dev agent or you can iterate over it the iteration part of the flow is written here so it basically goes in that loop QA dev QA dev until the QA agent passes everything and then it goes ahead to the merge sub agent.

So because Codex have now added in the ability to have sub-agents which they didn't have previously, this is all now handled with one command. Whereas previously I actually used to do this. So it would be dev 1237. Then I would have to spend about 20 minutes waiting for that to finish. Then I would do QA1237. And then once QA if QA failed, I would have to copy the response back up here.

And then eventually when QA passes and I would iterate over both of those then I would finally do merge 1237. So it's not like you need to do one agent at a time. I can have like 20 dev agents running which is exactly what I do. But now everything can just be handled because Codex has now added in the ability to have sub agents. It can all be handled just simply by running impl which again I've made that up.

So it's not going to be baked in on your machine. You actually have to set it up. If you want to set it up, you can just use warp and you can say, can you set up Zshrc to run impl issue number and create a worktree and pull in the issue and launch Codex over the CLI and then warp will just go ahead and and set that up on your machine for you. But also for any of you that were wanting to set up OpenClaw, if you struggled with it, you can literally just tell it.

The only thing by the way of this impl when you run like agents with sub agents on Codex, it does generally tend to spike CPU. I'm on like a M4 MacBook Pro here and the only time I'll ever see CPU go up to 100% is if I'm running a load of these workflows. Maybe like five or 10 implementation agents where they're launching sub agents will push CPU up to about 100%. The thing that's very interesting though, I'm not exactly sure why, but if you don't have it launching sub agents and I just launch directly work on this issue, I can run about 40 or 50 different agents at the same time, which was some of the things I was showing on my Instagram before without having any type of CPU problems.

So yeah, that is basically like my dev workflow. So again, everything from like this is the only part of the process and this is like a take on spec driven development that I've kind of iterated over, but this entire workflow is completely automated. Technically, you could also have the PRD agent actually implement all of this as well, but I do like to have a little bit more control of what gets launched and when.

Next up, I'm going to move over to my red team. This is my other like favorite thing to set up. I'm just going to get an official definition for you. A red team is an independent authorized group that kind of attacks your system. What I do is I tell the main agent, you are an external consultant. You're a team of external AI agents that have been hired to investigate and attack our system to look for bugs in the code, flaws in logic, things that aren't running as expected.

It could be security vulnerabilities. It could be overengineering of things that need to be simplified. The very important thing is to tell them your job is not to agree with things. You are external consultants. You're here to play devil's advocate and attack our system to show us the weaknesses and deliver a better outcome. So you would tell like the main agent again for the sake of it I'll just say so we launch Codex.

This is probably the only thing that I would actually look to put Codex in. I would do very high, extra high instead of just high reasoning and it's going to throw even more tokens at this. By the way, now is a really great time to do this because as I mentioned, Codex has a 2x limit on usage. So, you can just chuck a ton of compute at this. You are in charge of a team of external red team auditors.

Your job is to attack our code for flaws, bugs in logic, overengineered areas of the app that can be simplified and like security vulnerabilities I'll probably do separately. So again, I like to keep these really isolated. So this I would do more on like a kind of code quality. And again, as I mentioned, we're making like financial decisions. So it's really important that everything runs kind of flawlessly.

I want to design 10 to sorry to write 10 isolated GitHub issues that will be assigned to your team of red team to look for flaws and weaknesses. you can use MCP and run real world tests. So it's very important to give you that to to give them kind of that permission that they they've got like full autonomy not to make changes but they can can kind of write write their own test. So again it's probably going to because it's in high reasoning he he's probably going to want about 10 minutes to come back with stuff but that would basically come back with 10 issues.

I'll see if I can find if we come into red team, this is an epic for example, and these are a load of different investigations that got made. So, it's saying run a truth table and then this is an example of the report that actually came back. So, again, just to kind of go into my workflow, the project manager that we just set up, he'll come up with a load of instructions of what he wants the agents to do.

And then in parallel we can just kind of come across like this And then I just have the shortcut like investigate 1237 for example. Investigate 1238 or whatever the issues numbers are. And that will just go ahead and launch a load of agents to go and run investigations. Again it will make a worktree. That agent will then go and write up a full report. It will follow everything that was asked of it in the issue and it will write up a full report and it will leave that as a GitHub comment and then full report is then sent to GitHub and then either we can use this same project manager agent that we had here or we could do a PRD agent but tell it to review the report that we have back and again so this is going to come back with things like this needs to be improved you've got a bug here you need to fix this whatever it may be and then again we just assign those to dev agent QA agent merge agent the same as we have over here and they will work in parallel on fixing those things.

So I find this to be really good at having a bit of a self-improving and a hardening system again all driven by AI. Now another thing that I'll kind of talk to you about is I've been looking at setting up like different department most people are using cursor Claude Code um Codex for development work but now because of the addition of skills and everything there you can basically have a finance department so you can have your dev department which is what it's originally made for which is what we've just been going over for the past 30 minutes but then for example mine's an AI-driven hedge fund it can have a finance department.

Maybe you could also do legal department or or in marketing you could have different things like content department and then we could again using Claude Code or whatever it is you want. I can have like a quant researcher where I just say like you know your job is is and I' I have done this. So I've got a back testing engineer. I have a a data not a data engineer a it was a forensic data analyst and then I had a anyway this was like a new department that I set up.

So again it's all done through through Codex and through warp exactly like this. The only thing that I'm changing is is the prompting. So for example like the the quant researcher his job I've built out this like database which has 140 sorry 176,000 learnings using embeddings and it was basically everything that the AI agents have have learned I was putting into into this database a vector database so that the agents can actually retrieve from like a shared knowledge base.

So the researcher would basically pull in what are some things that the other agents have learned? What is research that we've done? What do we know to be true? What do we know to be false? And he would then use that to come up and design new back tests. The back testing engineer, his job was then to actually run those back tests because it's quite specific in how it needs to be set up.

The back testing engineer had a very specific set of instructions and then the forensic data analysis. His job was then to actually analyze the back tests. Look for like every for example if a back test runs over two years it would analyze day by day the good days the bad days what were the conditions on that particular day? What made it a good day? Was it down to one particular stock?

Was it that the market was up as a whole and therefore like our whole portfolio came came with it. It it goes through like a 13-stage audit that I wrote up as well. Also using machine learning, sklearn, which is scikitlearn is a Python module that I've been been working a lot with and really been enjoying. And then also doing things like rule mining. Again, I'm not going to go into it too much in this in this video.

But the the other point is is that that would have been like completely out of the rain the realm of possibilities for me a year ago. But now due to these tools I'm able to build such interesting like the things that I'm able to build have really changed. So they're the different AI departments that I've been setting up and then yeah just a little bit on machine learning but I think it it's quite technical so it's not really that interesting for most people.

But yeah if anyone is technical and playing around with this stuff I would really suggest you to look at this. And then the other question I keep getting is about OpenClaw or OpenClaw. I am using it. I'm not using it for any dev work. Everything, any development work that I do is entirely done through warp. And I pretty much spend my whole life just using those applications now. And I'm really happy with a very like minimal setup.

OpenClaw. Basically, I translated all of my notion stuff for it to understand that. So, it does a morning review and an evening review with me. It does a weekly review, a monthly review, a quarterly review, and a yearly review. But obviously I haven't actually done any of those yet because I've only been using it for about a week and a half. But I do like the fact that I can do my morning and evening review entirely through voice notes over Telegram.

I find to be really useful. So I've been much more consistent with that. And then I also have it prompt me kind of nudge me about habits or or like reminders and and things like that. So I've not given it access to everything, but I I was actually planning to build out something like similar to it before. not not nowhere near as advanced as like all the flexibility that's there, but specifically like an AI assistant for my notion system and I did make a video on that before on that I I built something with 11 Labs, but OpenClaw has like solved that problem far better than I did.

I also like that it's it's open source, so it's just going to see such rapid improvement because it's got quite an active developer ecosystem. But yeah, obviously like again, I don't want to repeat any stuff that that's like already out there. It's the thing that I hate most about this industry is like everyone's just posting videos about like this changes everything. The amount of times I see YouTube thumbnails with the term this this changes everything.

I I find it really annoying. So like on whatever security concerns and things like that, you can go and watch videos from other people. But yeah, I'm I'm literally just using it like a little bit here and there as like a proactive notion assistant. And I really like to do like the voice notes and sometimes I will use it as well as part of like the the PRD workflow that I showed you here.

Again, if I'm out and about and I can just send a voice note on on Telegram, I do really like that. Finally, just kind of closing thoughts. I think now is like just the absolute best time to be alive. For anyone working in like the techn technical space, now is like the best time to be a nerd that's ever been in history. I think we must see more innovation now every couple of months than 99% of humans that ever lived saw in their entire lifetime.

And the fact that we have like these huge companies in particular OpenAI and Anthropic just going head-to-head in trying to capture market share of these AI coding tools. I mean, it really benefits us. Like I said, Codex has 2x usage promotion on at the moment, so it it's just an incredible deal. And then obviously like Opus 4.6 out last week and then GPT-5.3 Codex they try to release on the same day.

So they're just kind of fighting each other but we as the consumers are the ones who really really benefit from that. I think I'll leave it there. This is like again it's mostly stuff that I've come up with myself in terms of the specific workflow but just pulling from different things. I do really like spec driven development and I think it is like really the way to go. This workflow does change quite a lot.

It's had 10 iterations maybe and I do change things like every week. I may look at putting something like open source on GitHub if people are interested, if there's enough interest or if people were willing to kind of help me maintain a repository or something because it's not just about like a one-time commitment. It's about constantly kind of updating things as I make changes and so on.

Or even perhaps I would do like a telegram group of like vibe coders for people to kind of come together and like share things or workflows that they're doing. I'm more than happy to do that of like sharing information together. Two final points. Number one, if people were interested in like a search tool for IG where you could basically do something like show me all the users that contain this specific keyword in the bio, for example, contain the word of or vibe coding and they have between 1,000 to 10,000 followers and they're based in London.

Would people be interested in a tool like that? Cuz we currently do have capacity to make something like that where you would basically be able to search the whole of IG. For example, you could use it for finding influencers or O of creators. Like we can segment by I mean it's a an AI generated guest but at like their specific job or based on like the link that they have in the bio, we can check if they do YouTube or of something.

So you would basically be able to filter every user on IG. If that's interesting to people, just let me know because it's something I've been considering looking at building out as as a SAS. And then the other one would be a trend finder basically of the whole of IG. So it would be analyzing every single post on IG that you can basically because it basically comes down to like how much data you're doing.

I know that a lot of people have now built out their own trend finders based off the videos that I was doing. But we would potentially be able to build something out that would cover billions of posts that you would get alerts for on Telegram for any videos that kind of match your criteria. Again, it can be based off of keywords, off of location, off of how big of an account it is, off of how viral the post is.

And you can use that content for AI repurposing, for having models copy it, for just reposting somewhere else, or just for like commenting on viral posts, which has been another interesting strategy as well. Anyway, I'll leave the video there. Thank you so much for watching.