Cursor's January 2026 planner–worker coordination experiment

Source: Wilson Lin, Scaling long-running autonomous coding, January 14, 2026; first-person engineering research account of internal experimental systems, not controlled productivity evidence. Early flat agent group used a shared status/task file with locks. Author says 20 agents effectively moved at throughput of 2–3 because most waited; optimistic concurrency removed lock friction but agents avoided hard end-to-end work. Later model has continuously planning agents define tasks, implementation workers complete them, and a judge decides whether to restart another cycle. An additional integrator role was dropped as another bottleneck. This is an architecture iteration on experimental tasks, not a head-to-head statistical comparison or measured human cost.

Outputs and boundaries. The prototype web browser ran nearly a week, wrote over a million lines and has linked public source; line count and screenshot are not product fitness or sustained maintenance. A separate three-week Solid→React migration had passed CI/early checks at publication but still required review; an unrelated video rendering optimization had been merged and was slated for production, not necessarily the same multi-agent planner/worker experiment. No reviewer minutes, cost per useful change, requester value or after-merge defects. One plausible adaptation to Dru’s manager-led practice is assign ownership of difficult outcomes and make dependencies/next decisions visible without shared-file lock ceremonies; do not translate vendor prototype’s generated-agent roles to replacement of human product or engineering managers. Coordination question; team adoption.