An orchestrator delegates. An implementer ships. An independent reviewer gates the finish line. You watch it happen — live, on your own hardware.
docker compose up — that's it.
The pattern
Each Claude is given exactly one job and exactly the context it needs — no more, no less. That's what makes the loop converge.
Plans · delegates · audits
Breaks the goal into small, well-scoped tasks. Every word it emits is the next instruction shipped verbatim to the implementer.
Reads · edits · runs · reports
Full workspace autonomy. Implements, runs tests, reports back concisely. Doesn't ask permission — gets it done.
Audits · gates the <<DONE>>
A fresh Claude with no conversation context. Reads the workspace cold and either confirms done or kicks fixes back to the implementer.
Built for trust
Every primitive built for the way real teams ship: durable, observable, and replayable.
Single docker compose. Your hardware, your keys, your data. No third-party SaaS in the loop.
Watch both agents think and ship in real time. Tool calls, diffs, tokens — all there.
Engine reboot? Loops pick back up exactly where they were. State is durable by default.
Point at any directory. Agents work inside it the way you would — git, files, shell.
A stop keyword plus a context-free reviewer means the loop only ends when the work is real.
Export any run as a clean markdown transcript. Great for PR descriptions and post-mortems.
Point at a repo, write a goal, hit run. Watch two agents ping each other until it's done.