Script your agents like you script your shell.
Work with agents in conversation. Script what repeats.
A shell for your agents
Your shell grew over years. Aliases for what you type every day, scripts for the chores you got tired of. Nobody designed it in one go.
Agents have none of that yet. Every session starts from zero: the same instructions, the same checks, the same corrections. Or you design a pipeline up front, and the agents disappear behind it.
coderoom is a shell for your agents. Work with them live, script what repeats.
On Monday you ask Ada to fix a failing test, run the suite, paste the failures, ask again. By Wednesday you write it down:
/loop @ada make the tests pass without weakening them /until /tests /max 3
Your tests decide when Ada is done. Three tries at most, every step visible in the room.
You write every script, and every script has limits. The room grows more capable while you stay in charge.
Your agents. Your rules.
Short iterations still matter
Software teams did not converge on short iterations because typing code was hard. They converged on them because short loops help clarify requirements, build shared understanding, and catch bad assumptions early. Agentic workflows should preserve that.
Get started
Download the latest release or build coderoom from source. You need Node.js with `npx` and a working Codex CLI setup. Support for other agent CLIs is planned.
Download the latest release
Choose the archive that matches your platform, extract it, and run `./coderoom`.
Build from source
You will also need Go. Then build, start a room, and script your first check.
$ make build
$ ./bin/coderoom
> /invite ada
> @ada implement a small change: ...
> /def tests /shell go test ./...
> /loop @ada make the tests pass without weakening them /until /tests /max 3
More output needs better scrutiny
One way to handle that volume is to split review by role. A review agent focuses on clarity and test coverage. An architecture agent focuses on boundaries and system shape. A security agent surfaces trust assumptions, unsafe defaults, or missing controls. A compliance agent flags where evidence, traceability, or audit expectations are thin.
What agents lack is engineering judgment and lived technical context. That is what the engineer brings. Used well, the combination is stronger than either alone: agents gather and surface the information; the engineer decides where to dig deeper and what deserves another pass.
coderoom insights
Writing on the engineering ideas that shape coderoom.
Mob Programming for One
Agents can generate useful material faster than an engineer can absorb it. This post uses mob programming as a model for structuring engineer-writer-reviewer loops that preserve understanding, scrutiny, and engineering …