Codex built me a full Memory Manager overnight using my new "Teams" strategy. Flawless first test. One prompt spun up 70 domain-expert subagents, organized into specialized teams, all working in parallel to one-shot the entire app while I slept. I've built over 200 custom subagent roles for this. Playbook... dropping soon. This is a massive unlock for parallelization. I'll be improving this and testing it on much larger projects.show more

How many tokens did this use? "Yes"

That's cool, I have a unified memory for all the agents' clients I use, and agents can access that with CLI or MCP.

That's reallycool. How's it work?

the first version was only for claude code 8 months ago, The new version is a client-server architecture for indexing and searching AI coding assistant conversations. A small client on each machine watches the chat history and sends only new or changed conversations to a central server. The server queues that data, a worker cleans/parses it, breaks it into useful chunks, and turns those chunks into vectors for semantic search (using a local embedding model). When you search, it combines keyword matching plus meaning-based matching, then returns the most relevant snippets through the dashboard or MCP tools. last addition i added a pi-web client so i can search directly from the web-app if i want it I will open-source it soon

You're ahead of your time sir. Well done

Shout out @CEOAlexColon for the idea. He's a good follow

I've been running something similar with specialized agent teams and the biggest lesson so far is that timeouts are non-negotiable. One stuck subagent at 3 AM can burn more tokens than the whole project was worth. Looking forward to the playbook.

If you have one that's stuck then you have another issue that you need to figure out and that's why auditing these processes is critical

100% agree. That's exactly what we found. We were trusting our agents too much early on and a stuck process was really just a symptom of a bigger problem... bad guardrails. Now every subagent gets mandatory timeouts, and we built an emergency restore system (we call it the Big Red Button) for when things go sideways. Auditing is how you go from 'it works sometimes' to 'it works reliably.'

Yeah once you build in the guardrails it works pretty well. But it definitely can take a lot of tinkering to get to that point. My goal is to do that tinkering so other people don't have to

Love it! Want to DM some ideas but can’t! Let’s pretend we are 10 again: “I really like your agents, would you like to be my friend”? ;)

Followed 🤝

sleeping while code gets written is the only startup perk I want

Amazing! My question, which I still don’t understand at all: How come you people build such brilliant ideas and give them away for free on GitHub?

Do it for the memes 🤪

lol, no really, I want to know. Or maybe your source of income is different and you're just helping people out.

I'm just building out a reputation and sharing the same free knowledge that was so generously shared with me. These ideas aren't totally novel. I just read and listen a lot. 🫡

Thanks bro!

Any time! Thank you 😊

I dunno if it’s cause it’s X but like you had a website with all your thoughts properly organised so I can review? I’ve been taking yo codex pretty seriously and would love to go to the next level.

I wish I had time, it's on my radar for sure. My adhd doesnt help 😆

Please try, I would love it. I just tried using sub-agents, and I got errors. Did you make a post how to set up sub-agents? I think you did 😭😭😭

/ experimental menu, make sure they're on

Codex is wonderful. GPT models are so good at coding.

They really are exceptional except... that UI lol

I would love to learn how to effectively work with so many parallel agents. Please do share some wisdom our way, if you can😅

Definitely click on the articles on my profile because I already have some tutorials out there!

Yep, found them! Thank you.

Super interested in seeing this playbook.

🤝

wait to see play book. Also working on porting old android apps to other platforms, any idea how to get full spec with old repos with no docs? The spec missing is a lot even with multiple agent exact and recheck loops

Start by having an agent look over all of the source code and talking it through

That’s an impressive example of parallelized development. Coordinating specialized subagents around domain expertise mirrors how high-performing human teams work — but at machine speed.

🔥🔥

Codex design is all the same? haha

Honestly though lol. I long for the day...

70 subagents for one app is the ultimate "agents as microservices" stress test. The real bottleneck is never the individual prompt—it’s the coordination tax as the graph grows. Would love to see how you’re handling state consistency across that many specialized workers without it becoming a message-passing nightmare.

Seventy coordinated agents overnight sounds like the first glimpse of synthetic org design.

The crazy thing is this is all still very Elementary in my opinion. We can go so much deeper

that's so cool, teams sound like a great strategy to organize those subagents, must be really helpful to have them working in parallel like

Seventy coordinated agents overnight sounds like the first glimpse of synthetic org design.

This is fascinating — the Memory Manager + team-based swarm is exactly the direction I've been researching from a neuroscience angle. I've been rebuilding the Complementary Learning System for AI: agents forming cognitive memory (facts, experiences, opinions) into mental models they can share across the swarm. Your "70 domain-expert subagents" setup is essentially what emergent collective intelligence looks like in practice.

how detailed was the plan? getting the right level of detail seems like the tricky part

It was a well thought out spec for sure but not the most complexity in the world. It was 1am and mostly just a proof of concept. Real torture testing to come. I need ideas though. Got any? Lol

DM'd you

this is actually insane i cant even make a normal chatbot work

Context, Context, context

How do you use your teams strategy? Do you have some tutorial? Thanks mate!

Not published yet, I'm still improving on it

Really looking forward to the playbook!

Thanks Eric! I'll try to finish soon

70 subagents with shared memory is where it gets interesting. The hard part isn't spawning agents — it's making sure agent 47 knows what agent 12 already figured out. That coordination layer is the real product.

that looks awesome

You look awesome

while you slept, your agents worked. that's the future right there

It truly is

Nice. Would love to learn more. Can you share and design doc etc ?

I'm going to put it all into a document and share it for sure I just need to organize my thoughts. But yeah definitely coming

have that strategy documented yet? 👀

Only in my peanut brain

That UI tho 🤮

@slopwareindy 😅😅😅

Yeah, wtb whatever this guy has been taking to be this efficient.

Haha I actually don't drink or take any chemicals except way too much caffeine

I am still trying to get my printer to work

nah thats hilarious

interesting

is the strategy available publicly for us to try out ? i'm interested

Soonish!

will be waiting , thanks

Awesome 👌🏿