Cursor details harness that let thousands of agents make most commits to a browser for a week
KyeGomezB · x · 2026-09-09
Cursor published a research post, Towards self-driving codebases, explaining the harness behind its long-running autonomous coding experiments.
The coordination scheme is elegant: workers are unaware of the larger system and never talk to each other. Each works on its own copy of the repo, then submits a single handoff document back to the planner that requested the task.
Key points:
- It began as a side project: prompting Opus 4.5 to plan and build a browser engine. Single-thread "keep going" prompting failed — the model lost track, declared premature success, but wrote good code in small pieces.
- Pivoting to a multi-agent harness, the system ran continuously for a week, with thousands of agents producing the vast majority of commits to a research web browser that was almost entirely runnable without human intervention.
- Part of the research is now open to some users to try.
More from coding & agent
- Tencent paper: evolving harder environments boosts terminal agents 18 points on Terminal-Bench — rohanpaul_ai · 2026-09-09
- Tencent paper: continuously harder task environments beat co-evolution, +8.6pp on Terminal-Bench — rohanpaul_ai · 2026-09-09
- RoboSim: open-source Rust robot-native simulation engine with physics, sensors and Python bindings — rsasaki0109 · 2026-09-09
- instructor 1.17 ships cache isolation fixes, Gemini retries, local PDF support — jxnlco · 2026-09-09
- Dev: Fable 5.1 Ultracode Workflows Deliver Best Quality, Codex Ultra a Waste of Tokens — jarrodwatts · 2026-09-09
- Put per-turn action schemas in the last user message to preserve prompt caching — Low_Bad_6585 · 2026-09-09