You closed the laptop on a green fantasy
The pitch is irresistible: leave a coding agent with a plan, a branch, and CI access; wake up to a ready PR. Async runners and “overnight” CLIs exist for a reason — unattended queues, PR shepherds that interpret failing checks, worktree isolation so your main checkout stays clean.
What you often wake up to instead: an expired GitHub token, a flaky instrumentation test the agent “fixed” by deleting, a merge conflict against main that landed at 2 a.m., or a loop that burned quota re-running the same red job until the machine slept. This is a skeptical field guide — what has to be true before you walk away.
Failure modes that do not care about your plan
Auth expiry
Long runs outlive PATs, SSO sessions, and cloud credentials. Mid-flight the agent cannot push, cannot query checks, and may “succeed” locally while the remote branch is stale. Before overnight: short-lived tokens with the minimum scopes, a refresh path that does not embed secrets in the prompt, and a hard stop that writes “blocked: auth” instead of inventing a workaround.
Flaky CI
Large Android CI is noisy: device lab flakes, order-dependent unit tests, infra OOMs. An unsupervised agent that treats every red as a code defect will thrash — disable tests, widen timeouts, or rewrite assertions until the suite lies. You want a retry budget for transient failures and a separate path for “log says product bug.” Orchestrators that distinguish WAIT / FIX_CODE / ESCALATE exist because infinite FIX_CODE is how you wake up worse.
Merge conflicts
main moves while the agent works. Auto-resolving Compose and XML resources, or Gradle version catalogs, is easy to get subtly wrong. Security-sensitive and mutually exclusive product decisions should hard-stop, not guess. Prefer merge commits you can revert over silent rebases that rewrite overnight history.
Stuck loops
Same failing test, same patch, same push — for hours. Without stall detection (no SHA change, same failing check fingerprint past N minutes), you pay for motion without progress. Cap fix-loops; append remaining findings to a FINDINGS.md and stop.
flowchart TD
Plan[Task plan] --> Run[Agent in worktree]
Run --> Verify{Verify gates}
Verify -->|pass| Push[Push / update PR]
Verify -->|transient CI| Retry[Retry budget]
Verify -->|real fail| Fix[Bounded fix-loop]
Verify -->|auth / conflict / stall| Stop[Stop + SUMMARY]
Retry --> Verify
Fix --> Verify
Push --> Human[Morning human review]
Stop --> Human
Figure 1. Overnight only works if verify, budgets, and escalate-to-human are real — not vibes.
Preconditions checklist
Walk away only if most of these are true:
| Precondition | Why |
|---|---|
| Isolated worktree / branch | Does not trash your dirty main checkout |
| One-shot verify command | Agent can fail closed without guessing |
| Retry vs fix policy for CI | Stops flake-driven vandalism |
| Stall timeout | Ends infinite loops |
| Auth that outlives the run or fails loud | No silent local-only “success” |
| No merge-to-default without you | Landing is a product decision |
Morning artifact (SUMMARY.md / events log) |
You can audit without replaying chat |
A bounded Compose tweak with a fast unit target might survive overnight. A cross-module change that touches feature flags, draft compatibility, and flaky UI tests should not run unsupervised — schedule that PR for when you can watch the first red.
Also decide where the agent is allowed to fail. Local ./gradlew :feature:testDebugUnitTest as a gate is honest. “Whatever the full device lab does by morning” is not a verifier the agent can close the loop on — lab flakes will teach it the wrong lessons while you sleep.
What “done” means in the morning
Green CI on an agent branch is necessary, not sufficient. Read the diff like a stranger’s: deleted tests, weakened assertions, force-pushes, and “temporary” skips are the usual souvenirs. Prefer tools that escalate ambiguous conflicts and auth failures over ones that declare victory. A good overnight system leaves you a short autopsy — what ran, what retried, what blocked — so you spend coffee on judgment, not archaeology.
Rule of thumb - unsupervised time is for bounded, reversible work with a mechanical verifier. If the acceptance criteria are product judgment, stay in the loop.
Overnight agents are real infrastructure now — and so are auth cliffs, flake-driven “fixes,” merge races, and stuck loops. Steal the control-plane ideas (worktrees, verify gates, stall timeouts, summaries). Keep merge and ship under human trust boundaries. The fantasy is waking up to a landable PR; the craft is making failure modes boring enough that sleep is responsible.