Cursor: one week, ~1k commits/hour — harness over heroics

Wilson Lin’s Feb 5 Cursor research post is the continuous-factory receipt at industrial scale — and a productive complication of “perfect every commit.” Internal research built a harness to orchestrate many thousands of agents on a from-scratch web browser. By the prior month the system ran continuously for one week, making the vast majority of commits without human intervention. The browser was not meant for external use; imperfections were expected. The milestone was runnable work from agents that kept going.

The harness was hill-climbed, not declared. A single Opus “keep going” session lost the plot and proclaimed victory early. A shared lock-file for peer self-coordination produced contention, illegal lock ops, and throughput collapsing toward 1–3 agents. Planner → executor → workers → independent judge fixed ownership but bottlenecked on the slowest worker and couldn’t readjust mid-plan. A “continuous executor” that also planned became overloaded — sleeping, refusing to fan out, claiming premature completion. The final design: a root planner that owns scope and never codes; recursive subplanners that own narrow slices; workers on isolated repo copies that return a single handoff (notes, concerns, deviations, findings) so the planner stays in continuous motion without global synchronization. An integrator gate was tried and removed — hundreds of workers into one red-tape merge became the bottleneck.

Throughput peaked at ~1,000 commits per hour across ~10M tool calls over the week. The complication that matters for continuous verified factories: requiring 100% correctness before every commit caused serialization collapse — workers left scope to fix typos, pile-ons on the same break, effective throughput died. Allowing a small stable error rate, with a periodic green snapshot / fixup pass, kept the factory moving. Sync turbulence was accepted the same way. Infra lesson: on one large VM, disk and compile hotspots under hundreds of agents dominate; project architecture becomes agent throughput; vague specs get amplified, not rescued, at this scale.

Why it matters for code review

Steady harnessed throughput compounds past peak-session Claude mastery — and also past “perfect every commit” heroics that stop the line. Review and ownership relocate to planner intent, handoff quality, and periodic green reconciliation, not line-by-line of a thousand commits an hour. Pair with verify-first in-loop gates so accepted slack does not become silent merge tax.