teardown · Peter Steinberger
3–8 Codex agents, one folder, one 800-line AGENTS.md
No worktrees, no subagents, almost no MCPs. Just a grid of Codex terminals sharing one folder and one long rules file.
01
The setup
Peter Steinberger describes running 3 to 8 Codex agents at once in a 3×3 terminal grid, mostly in the same folder of a TypeScript app of about 300k lines. Each agent makes atomic git commits. One AGENTS.md of about 800 lines is symlinked to CLAUDE.md so every harness reads the same rules. tmux handles background tasks, and slash commands like /commit, /automerge, /massageprs and /review handle the routine work.
02
What he cut
Worktrees and a PR per change went because they were slower and he runs one dev server. Subagents were replaced by separate terminal windows. Most MCPs were replaced by CLIs: gh instead of the GitHub MCP, which he says saved 23k tokens. Playwright MCP gave way to chrome-devtools-mcp.
03
Our read
This setup is mostly a bet on supervision over isolation. Many agents in one folder only works if each commits small, atomic changes, and the 800-line rules file does the coordinating that worktrees would otherwise do.
04
What to steal
Have parallel agents make small atomic commits in one folder, and replace any MCP that has a CLI equivalent.
05
Caveats
The model named (gpt-5-codex) is older than today's frontier; the source has no date in our record, so check it before copying model choices.
Facts from https://steipete.me/posts/just-talk-to-it. "Our read" is Agent Setups' analysis, not the author's words.
next teardown →
Four tools. No MCP. No sub-agents. It ships.