Overnight agents that open draft PRs: the harness is the product

Pickuma (May 20; also pickuma.com) strips the overnight agent pitch to a state machine: ingest → plan → execute → verify → package. The anti-pattern is collapsing plan/execute/verify into one “here is the repo, output the diff” prompt — fine for three-line fixes, fatal for anything larger. Stage four is the product: feed failing test/typecheck/lint text back into the next edit, bound retries (three is their ceiling), and if still red stop, open a draft PR, and log the failure instead of burning tokens until morning.

GitHub wiring that kept them safe in real repos: one predictable branch per task (agent/<issue>-…) so reruns are idempotent; draft PR assigned to a human; agent-generated label as provenance; CI on the bot PR exactly as on a human’s; fine-grained PAT scoped to one repo; execute in a container, never the laptop; branch protection on main so the worst case is a bad PR. Trigger is cron or agent-ready issue label. Honest scope: mechanical migrations, coverage, codemods, config tidy — “done” objectively checkable. It stalls on ambiguous product judgment and on thin test suites, where confidence is not a gate. Two numbers decide keep-or-kill: cost per run (hard dollar/token ceiling) and PR acceptance rate without substantial rewrite.

Why it matters for code review

FOMO that nothing moved overnight is cheaper than waking to rubber-stamped green. Continuous harnessed momentum beats bursty Claude mastery only when the night ends at a reviewable draft with a real verify loop — not a merge, not a self-declared success. Treat every overnight diff as untrusted until CI is green and a human has read it; the morning review is the merge criteria, and the agent is the night shift that handles only the boring, checkable parts.