Let it run for days — only if verify is a gate, not a vibe

Pratik Patel (September 22) left an agent on a real UI bug for three days — sign in, hit the broken screen, reproduce, read the trace, open a PR, watch CI, adjust — and woke up to a stack of dead ends plus a running log still inside the product. The punchline is not the clock. Letting it run is easy; making the run bounded, reversible, and inspectable is the skill. A pasted “fix onboarding” walkaway is an unsupervised intern with your credentials, not a multi-day agent.

He treats a day-long goal as a product: a stop condition stated in the product (“CI green and the repro fails to reproduce”), a named blast radius (accounts, environments, write actions), and a log that survives the session when tokens expire and context fills. The working loop is deliberately boring — reproduce → trace → fix → verify on the same scripted path — because chat-only agents collapse into confident summaries and skip the gate. If it cannot re-enter the UI and fail the old repro, it is drafting. Before walking away: named owner, environment seams, spend/time/tool-call budget, handoff format (repro, hypothesis history, current diff, next action), and a one-sentence kill switch. Days surface quiet failures short sessions hide — context rot, false fixes (widened try/catch, deleted flaky asserts), scope creep, tool thrash — and he treats overnight as a deliberate mode: can the agent still prove progress with a fresh repro if the chat dies? Screenshots and trace IDs beat status sentences.

Why it matters for code review

Always-on momentum only compounds past peak Claude sessions when the harness owns verify-on-the-real-path and the kill switch — not when someone stretches a chat. Patel’s checklist is the FOMO antidote that still refuses theater: buy wall-clock time against a defined ticket, wake to candidacy with receipts, keep humans on irreversible promote.