Scripting your agent chats: how I stopped copying text between my models

I love working with models in chat. But I ended up with one model building and others reviewing code, architecture and security, and I was copying text between them all day. After a while I noticed my decisions got worse: my attention went into coordinating models instead of thinking.

So I built coderoom, a shell for my agents. They share one room and one git workspace, I talk to them like in a chat, and I script what repeats.

The talk is a live demo. Two agents fix a failing test: Ada builds, Turing reviews, and nobody copies and pastes. Then I turn the repeating part into one line:

/loop @ada make the tests pass without weakening them /until /tests /max 3

Then a look under the hood, and three things other builders can reuse without coderoom:

• Build on the agent CLI, not around it. Each agent runs as a codex app-server process over JSON-RPC, so you get everything the CLI already does and spend your time on what only you can build.
• Record and replay. Agents answer differently on every run, but tests must run deterministically. Record real sessions and replay them in CI: no model calls, and a failing test is your bug, not a new Codex version.
• Don’t let the agent decide when it’s done. Let an external check decide, like your test suite or static and dynamic analysis, with a hard limit on tries.

coderoom is open source: github.com/roomscript/coderoom. More on the coderoom page.