Is your project doing work overnight yet?
Peak Claude sessions are calendar-bounded. A harness that keeps modest verified momentum through the night compounds past glory bursts — but only if morning inventory is judged, not just generated.
Peak-session mastery is real. You can still watch a coding agent crush a hard slice in a glory window and feel like the factory is winning. The better question is quieter: is your project doing work overnight yet?
The itch behind that question is familiar to anyone living in agents. Leaving a lab or a repo idle used to be normal downtime. Now every night without agents working for you is progress left on the table — FOMO as opportunity cost. The productive answer is not a longer chat. It is infrastructure that keeps moving after you close the laptop: detachable runtimes, progress files, oracles, and stop conditions that refuse a fake “done.”
Glory sessions vs compounding nights
A watched Claude Code session is still bounded by the human calendar. A regular week is roughly forty judgment hours; a grind stretches toward seventy-two; an agent’s week is a hundred sixty-eight clock hours — and agents do not take holidays. Those hours are not interchangeable (late grind ≠ morning judgment), which is exactly why the harness matters: it is how you spend the calendar without minting review debt.
The field is full of receipts that continuous motion compounds past heroics. Cursor’s self-driving harness ran about a week with ~1,000 commits per hour once planners, workers, and handoffs replaced a single “keep going” session (primary). Stripe’s Minions pattern ships 1,300+ fully agent-produced PRs a week by interleaving agent nodes with deterministic lint/push/CI and bounding repair loops (primary). Auto mode’s production unlock is continuity itself — 9× longer between interrupts, with Nuro’s 10pm–5am hill-climb waking to three PRs (primary). Smaller shops rhyme: Ralph loops with stop hooks and machine-checkable done (primary); Nightwatch fleets that plan by day and ship overnight (primary); headless Decode on Modal that turns cron into morning PR inventory (primary); Checkly’s harness-first overnight rewrite; and the overnight agent pattern written as queue → isolated exec → watchdog → morning triage (primary).
Picture a burn-down of remaining work. A harnessed line keeps falling through the grey overnight bands. A bursty mastery line drops hard in glory sessions, then flats — or ticks up as unverified debt arrives with the morning coffee.
FOMO without a judge is just token theater
Overnight motion only redeems FOMO when delivery stops at verified inventory. Cognition’s own shift is the industry tell: more agent sessions are now triggered asynchronously than by a human in the chair, and generation stopped being scarce — verification is (primary). Long-running agents need harnesses, not hope: task contracts, scoped workspaces, tool policy, checkpoints, verification gates, cost/time budgets, and a final receipt (primary). Cherny’s operating rule is the same invoice at Anthropic scale: production Claude code needs a higher bar than human code, paid as continuous lint, tests, e2e, daily fuzzers, and automated review (primary).
Without that stack, overnight is expensive theater. One community overnight run fixed twelve design findings and passed an independent rebuild — then burned 158M tokens, ~95% re-reads (primary). Verify-after still invoices the Audit Tax: ~90 seconds to draft a PR, ~90 minutes to trust it, with AI PRs taking 4.6× longer to review (primary). Verify-first harnesses exist because 46% of agent fix PRs never merge when gates sit outside the loop (primary). And green CI is not a morning win condition: one autonomous Claude Code session merged ~24 green PRs and left ≥9 live regressions the same evening (primary). Cursor’s week-long harness even found that 100% correctness before every commit serialized the factory to death — modest accepted slack plus periodic green reconciliation beat perfectionist heroics.
Dual-LLM Night Shift patterns, independent rebuild judges, and blueprint-bounded CI are the same idea: overnight throughput compounds only when an unskippable judge exists before the human is the regression detector.
Why it matters for code review
Review relocates. The scarce morning job is not replaying the chat; it is picking up a verified inventory — PRs with oracles, traces, and gates that already failed closed — and owning the harness that keeps that inventory honest across the night. Peak Claude mastery still wins afternoon slices. The projects that pull ahead are the ones whose remaining-work line keeps falling while everyone else is asleep.