Overnight coding agents: attention-decoupled runs with a ready-or-not verdict

Ralph Workflow (May 30) names the overnight coding agent pattern as an attention-budget move, not a speed trick: a 4-hour unattended run should cost ~5 minutes of human focus (spec + morning review), not four hours of babysitting. The anti-pattern is being the runtime — prompt, nudge, re-review in a tight chat loop. The pattern inverts that: write a spec with success criteria (not a one-line prompt), launch, genuinely disconnect, and return to a git branch with commits, a build/test verdict, a checklist derived from the spec, and a diff you can git show.

They draw a hard line: a generator with a sleep timer is not an overnight agent. The run must be time-decoupled, structured (plan → implement → verify → fix → verify), verifiable (proof checks ran and passed or failed), and reviewable (repo artifacts, not a chat transcript). Morning triage starts from a ready-or-not signal within about a minute. Fit is bounded refactors, specced features, test generation, docs, dependency upgrades with automated tests; bad fit is exploratory work, undefined legacy behavior, threat-model-heavy security, or “I’ll know it when I see it.” The economics claim is familiar factory math: evening attention spent on specs and review, unattended hours on the inner loop — judgment at the bookends, not mid-execution.

Why it matters for code review

Continuous factories beat peak-session heroics when overnight work arrives as reviewable evidence instead of a transcript you have to reconstruct. Ralph’s checklist — structured loop, machine verdict, real diff — is the FOMO antidote that doesn’t hand the merge button to the agent: you still decide merge / request changes / keep the good parts, but you stop paying attention tax for every verify-fix cycle.