06 / IN THE LOOP · Available
Done is an exit code. Not a declaration.
A carry-on harness that keeps agents looping — orient, claim, act, verify — until exit codes prove done.
Agent Playbook is our open-source harness for long-running agent work (github.com/riverho/agent-playbook, MIT). Drop one folder into any repo and an agent can loop without friction: orient on a master file, claim a task, do the work, prove it with executable checks, record, and report.
- Enforced done via executable acceptance checks
- Anchors and checkpoints for long-running loops
- Complete cycle: loop engine, orchestrator, flows
- The problem
- Agents lose the thread between sessions, drift from process, and declare work done without proof. Prompts alone do not survive long runs: context compacts, attention wanders, and finished means whatever the last message claimed.
- Our contribution
- Designed and built the complete loop cycle. The `pb` CLI carries the whole rhythm: status, next with claim, record, report, validate, anchor, checkpoint, loop epochs with learn, background run with ps and stop, plus scaffold for carry-on installs. The verification gate refuses a done record when any acceptance check fails. Anchoring re-injects the North Star every turn, checkpoints detect drift and snapshot cold-resume state, and the mode-agnostic orchestrator composes monitor modes into flows with explicit artifact handoffs and capability-gap proposals.
- How it works
- Every iteration runs orient, select, act, verify, record, report, repeat. Tasks carry acceptance checks as shell commands; recording done re-runs them and exits non-zero on failure, with skips stamped visibly on the record. Loop epochs stamp each entry with a loop id, quarantine contaminated runs, and require a recorded lesson before the next loop opens. The orchestrator heartbeat scaffolds a backlog from each mode's own descriptor, drains it to done, and surfaces errors or missing capabilities through append-only logs.
- Outcome
- The engine is published open source at v0.3.6 and runs here too: this site's own `.agents-playbook` install carries the CMSA North Star through the same loop. No client results are claimed for it.
No published sources for this yet.
Or open Contact to start an inquiry.
