productagents
rbs agent "fix CHK-41" --mode accept_edits
Agents that work like engineers. Build, test, open a review.
Run coding agents on branch servers with the tools people use: the build graph, Atlas, tests, stacks. Every run ends in a change request, gated like anyone else’s.
how you use it
Coding agents, in order.
01
Hand it the task, in the branch
From the editor’s agent panel or the branch terminal, on the server where the code already builds. The run’s mode decides what it may do on its own: accept edits, ask for commands, plan only, or everything. A headless run denies what it is not allowed rather than hanging on a prompt.
02
It reads the graph and runs the real tests
Forty-five tools: file edits, search, commands, build and test targets, dependency queries, Atlas, stacks, tasks, subagents and teams. It runs the same test target you would, with the same coverage floor, on its own branch server.
03
Every step is revertable
Checkpoints are Merkle-tree snapshots of the workspace: unchanged subtrees are shared, so a diff across a monorepo is cheap and a revert is exact. Accept and reject feedback becomes taste signals for naming, structure and verbosity on the next run.
04
It ends in a change request, like anyone else
The run opens a change request on the stack. The same branch rules apply: required checks at the live head, approvals that go stale on the next push. Bring your own model with Anthropic, OpenAI, Google or xAI keys, or the hosted relay with per-organization metering; bring your own agent, since Claude Code and Codex run as engines inside the harness.
not built yet: an organization-wide agents console
The full reference lives in the docs.
works with
Open a branch and try it.
Boots in seconds. Your team joins with a link.