What if AI could manage an entire software engineering team, including reviewing its own work with a competing AI?

ChatGPT Feature requests
, , ,
1

What if AI could manage an entire software engineering team, including reviewing its own work with a competing AI?

This is no longer theoretical. I’m doing it right now.

I recorded a live demonstration of Claude Opus 5.5 working alongside OpenAI’s GPT-6.1 Sol to build a large-scale, multi-role application for a client.

And here’s the interesting part: I’m not manually coordinating these AI agents anymore. They’re communicating and coordinating with each other.

:robot: What’s happening in the video?

Think of a traditional software engineering team:

:man_technologist: Engineer 1: Working on Feature A
:man_technologist: Engineer 2: Working on Feature B
:man_office_worker: Engineering Manager: Coordinating both engineers, resolving dependencies and escalating decisions.
:magnifying_glass_tilted_left: Independent Reviewer: Reviewing designs and challenging technical decisions.

Now replace all four with AI.

That’s essentially what I’ve built.

Two independent Claude sessions are working on separate development tracks (we call them lanes).

A third Claude session acts as the Orchestrator, coordinating their work, resolving dependencies and bringing important decisions to me.

When independent technical judgment is needed, Claude invokes OpenAI Codex running GPT-6.1 Sol at High reasoning effort to review designs and challenge assumptions.

:high_voltage: How do these AI agents communicate?

It’s surprisingly simple.

The Claude sessions exchange messages directly through a cross-session messaging mechanism. They also share documents and technical decisions through structured files.

When one agent sends a message, another can receive it, process it and respond.

No human copying prompts between chat windows.

And no, this isn’t an out-of-the-box Claude feature.

This orchestration system was built collaboratively by Max (me), Goku (Codex) and Naruto (Claude).

:money_bag: But there’s an important catch.

Every AI interaction consumes tokens, which represent computational cost.

Even a simple “Received, thanks!” between agents can trigger processing of a massive conversation history.

We discovered that excessive communication and unnecessary reviews can make autonomous AI development surprisingly expensive.

So we’re optimizing not just how AI writes code, but how AI agents collaborate efficiently.

:bullseye: My biggest takeaway?

The next leap in AI-assisted software engineering isn’t simply better code generation.

It’s AI systems that can plan, delegate, collaborate, challenge each other and know when human judgment is necessary.

We’re moving from AI as a coding assistant toward AI as an engineering organization.

This video is a glimpse of that future, running on my MacBook today.

:movie_camera: Watch the live demonstration: [https://www.youtube.com/watch?v=16essfEOc3U\]

#ArtificialIntelligence #ClaudeAI openai Codex #AgenticAI #SoftwareEngineering #AICoding #MultiAgentSystems

2

Request to moderators: Extend the discussion window

This post shares a working implementation of multi-agent AI orchestration involving Claude and OpenAI Codex.

I’d genuinely like to engage with developers, answer technical questions, and exchange architectural insights.

A 24-hour discussion window makes that difficult, particularly for members across different time zones.

Could this topic remain open for at least 7–14 days to encourage meaningful technical discussion?

Thanks!

3

The competing AI reviewer is the part I find interesting. Having an agent build something and another one challenge its decisions sounds more useful than just throwing multiple agents at the same task.

But I’d still want a human to review the important decisions. Two AIs can make the same wrong assumption and end up agreeing with each other.

I’m curious how well this would work on an actual project with a messy codebase, changing requirements, and all the little decisions that aren’t documented anywhere. That’s where I’d really want to see this tested.

4

To be honest, i have the exact same question that when this setup has to go against a legacy code base then how well will it work? The legacy code base will have too many files hence too large to process and it will have zero knowledge base around how code is organized or audit trails. One comforting fact is that Claude has already written around 95k lines of code so current code base is NOT small by any standards. May be its a medium size project of a legacy behemoth.

My approach right now is simple, AI writes better code 100% times and AI does code/design reviews faster than humans 100% of the time and better around 80% of the time. The problem that i see with automated reviews is that both LLM agents start to chase perfection which led to 11 full loop of review and response once for me. I know how to stop that efficiently but i dont feel confident to share that right now :slight_smile:

one more thing is that this way there is NO collaboration as i am the single person working as EM, PM, TPM, Architect and 5-6 SDE(s). How do we get it up and running for a team is a different mountain to be conquered in itself? Or may be we dont need a team in future hence this gets short circuited into an auto close