Operating Model

High throughput comes from short, durable loops.

The goal is not more agents. The goal is more reviewable work safely merged.

Durable workers checkpoint state, retry safely, emit useful logs, and stop when an operation needs human judgment. Agentic engineering needs the same shape: branch, change, verify, commit, hand off, and escalate before risky work.

The Loop

1. write a clear task brief
2. assign a narrow surface
3. branch from latest main
4. make one reviewable change
5. verify the touched surface
6. commit atomically
7. open a focused PR
8. return a review pack
9. merge only after human approval

Durable Work Mapping

Durable worker concept	Agentic engineering equivalent
execution context	branch from latest main
checkpoint	atomic commit
health check	targeted verification
execution log	review pack
retry safety	dry-run, idempotency, and rollback notes
circuit breaker	stop-before-touching policy
handoff queue	pull request

Agent Task Brief

Goal:
Fix failed checkout retry state.

Scope:
- checkout UI
- checkout API retry handling
- focused tests only

Do not touch:
- billing provider config
- migrations
- pricing
- auth

Verification:
- targeted checkout tests
- typecheck
- manual failed-payment smoke checklist

Parallel Work Rules

• split by ownership boundary
• avoid multiple agents editing the same files
• prefer one PR per intent
• keep shared config changes separate
• rebase before PR
• stop when main moves across your touched surface

Review Queue Hygiene

Healthy queues have small PRs, explicit verification, rollback notes, clear human decisions, and no mystery generated files.

Unhealthy queues have long-running branches, mixed unrelated changes, risky data changes without dry-run, vague summaries, and PRs opened from stale main.