What Cursor learned about running many agents

Claim

Cursor's published lessons are the most concrete public record of what orchestration shapes work. Flat self-coordinating swarms collapse (20 agents gave the throughput of 2 to 3). Recursive planners, workers that do not coordinate, and a judge agent scale to hundreds of concurrent workers. An "integrator" role was removed as a bottleneck. Projects (September 2026) productises this as a coordinator that only delegates, keeps shared context files synced, and subscribes to Slack channels, schedules and PRs.

Evidence

• Scaling agents (2026-01-14): FastRender browser, about one week, over 1M lines of Rust, hundreds of concurrent workers on one branch (peak about 2,000 agents, about 30k commits per Willison). GPT-5.2 judged better at sustained autonomy; "Opus 4.5 tends to stop earlier and take shortcuts". "The harness and models matter, but the prompts matter more." Open problems: planners should wake on task completion; periodic fresh starts against drift.
• Long-running agents preview (2026-02-12): plan then approve, per-model custom harnesses, multiple agents verifying against the plan, targeting "drift" and "partial completion".
• Third era (2026-02-26): over a third of merged Cursor PRs from autonomous cloud agents; "build the factory that creates their software".
• Projects (2026-09-10): a persistent coordinator that never writes code and delegates to "thousands of subagents"; "because it delegates rather than executes, it is never blocked"; cloud by default, spawns a local agent when something must run on the user's machine; Subscriptions watch a Slack channel, run on a schedule, or follow all your PRs and fix CI. A design-system "gardening" project adds a lint rule when the same mistake appears twice. Projects-primary users merge 6x more PRs (vendor number).
• Per-model harness tuning (2025-12-04): for Codex, tools renamed to shell-like names; dropping reasoning items between tool calls cost 30% on Cursor Bench; a "conserve tokens" instruction made Codex refuse ambitious exploration and was removed.

Open questions

• Whether Projects' "shared context files synced across machines" is a plain file sync or a structured store.

Sources

• https://cursor.com/blog/scaling-agents — blog, Tue Jan 13
• https://cursor.com/blog/long-running-agents — blog, Wed Feb 11
• https://cursor.com/blog/third-era — blog, Wed Feb 25
• https://cursor.com/blog/projects — blog, Wed Sep 09
• https://cursor.com/blog/codex-model-harness — blog, Wed Dec 03
• https://simonwillison.net/2026/Jan/23/ — blog, Thu Jan 22

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.