One Claude Code session can run others without babysitting

A coordinator session spawns worker sessions of Claude Code, then re-executes their claimed commands and reads the diffs itself, according to a writeup published September 19. That matters for anyone running multiple agent sessions against one codebase, where a worker's summary can claim success while the diff shows nothing changed. The eight-step loop treats a worker's report as unverified until the coordinator confirms it independently.

The loop runs in eight steps. The coordinator pulls an unblocked task from the board. It runs the verifier first and requires it to fail, a step the writeup calls a RED GATE. It delegates the task to a worker session. Then it re-executes the commands the worker claims to have run and reads the resulting diffs itself, rather than trusting the worker's summary.

The coordinator itself never writes code. It spawns workers through cmux and keeps all task state on a Plan Desk MCP board instead of conversation context, per asyncdot's writeup.

Plenty of teams already run parallel agent sessions without a shared board or a coordinator role. What's different is refusing to trust a worker's self-report, and re-deriving the result independently instead. Any multi-agent setup can borrow that discipline without adopting cmux or a Plan Desk.

The writeup names two mechanics that transfer to other setups. One: verify a spawned session started by checking process timestamps, not by assuming the spawn call succeeded. Two: in a shared repo, commit with git commit -- <paths>. A bare git commit lets one session sweep up and commit files another session has staged.

The writeup doesn't say how the approach holds up past two or three concurrent workers. It also doesn't address what happens when the verifier itself has a bug. Anyone running more than a handful of sessions against one repo is the test case the setup hasn't faced yet.