abhishek.it
Back to posts
aiagentscodexopen-source

Building Native Agent Team Mode for Codex

How I added Claude Code-style team coordination to Codex — one lead agent spawning and coordinating multiple teammates in parallel using a shared filesystem task board.

I've been hacking on Codex (OpenAI's open-source CLI agent) and one thing missing was the ability to run multiple agents in parallel, coordinated by a single lead. Claude Code already does this — lead agent spawns teammates, creates a task board, teammates poll for work, claim tasks, write results back. Filesystem-based protocol, no orchestration layer. So I built the same thing into Codex.

How it works

The lead creates a team and a shared task list on disk. It breaks down the work, creates tasks, spawns teammate agents. Each teammate polls the task board, claims a task via atomic file write (no race conditions), does the work, writes results back, checks for more.

The lead reads results straight from the board. No message passing, no pub/sub, no database. Just files.

Why bother?

GPT-5.3 Codex pricing is way cheaper than Claude Opus. Run 3-5 agents in parallel doing research, writing code, running tests — that difference adds up. Even for silly test cases like "write haiku about nature, technology, and time," you get three agents working simultaneously instead of one grinding through them sequentially.

The coordination pattern

Lead Agent
  ├── Creates team + task board (filesystem)
  ├── Breaks work into tasks
  ├── Spawns N teammate agents
  │
  ├── Teammate 1: polls → claims → works → writes result
  ├── Teammate 2: polls → claims → works → writes result
  └── Teammate 3: polls → claims → works → writes result

Each teammate is its own process. They never talk to each other — coordination is just reading and writing files. The lead monitors progress by checking task statuses on disk.

Shared store, not message passing

Someone asked if the lead handles state sync or if there's a shared store. Shared store, per team. Lead creates tasks, agents poll the board, claim work, write results. Lead reads results directly. No relaying. Agents only use team_message to ask the lead questions, not to send results.

Same filesystem protocol as Claude Code. I added it to Codex because the pattern is simple and GPT-5.3 pricing makes running a whole team of agents cheap.

Will share the fork soon.