Throw a task in. The room engineers its loop.
No fixed pipeline — the loop is synthesized for this task: its gate, its critics, its revisions, and whether a human has to sign.
Loop engineering, run by a band of agents
Every coding-agent tool runs the same fixed loop for every task — a CSS tweak and a database migration get the identical Planner→Coder→Reviewer pipeline. When the loop's exit condition is too weak, confident-but-wrong code ships.
Ouroboros reads a task and synthesizes the loop that task needs — the checks that gate it, the critics it recruits, whether a human must sign — then a band of agents runs it on one audit trail.
1EngineerLoopArchitect designs a LoopSpec from the task's kind + surfaces.
2RecruitRivalReviewer pulls in specialists on demand via
band_add_participant.3RunCodeAuthor writes, QAEngineer runs the real checks in a sandbox, the loop bounces.
4Gate & shipA human signs only when the surface is high-stakes. Ships green or hands a verified failure.
The loop is real: checks execute generated Python in a sandboxed subprocess. A buggy revision genuinely fails with an AssertionError; only the fix passes. Open any real-QA card to read the actual code, fix, and traceback.
run it on your own repo
pipx run ouroboros trykeyless · runs the real loop on a bundled buggy repo
python demo.py "your task" --kind feature --touches auth,db