Cursor's planners, workers, and judge

Roles

This is Cursor's account of its own research, not an independent write-up. The post describes role types and says hundreds of workers run concurrently. It does not name individual agents. An integrator role was tried and removed.

• Planner

  Owns. Explores the codebase and creates tasks. Planners can spawn sub-planners for specific areas.

  Constraints specific to this role

  • To be added.

• Worker

  Owns. Picks up an assigned task, finishes it, and pushes the changes. Workers do not coordinate with other workers.

  Constraints specific to this role

  • Coordinate with other workers. The post states that as the role design.

• Judge

  Owns. At the end of a cycle, decides whether to continue. The next iteration then starts fresh.

  Constraints specific to this role

  • To be added.

Routines

• Plan, work, judge

  Each cycle

  Planners create tasks, workers complete them, and the judge decides whether to run another cycle from a fresh start.

Guardrails

• Workers do not coordinate with each other and do not handle the big picture.
• Each new iteration starts fresh. The post says periodic fresh starts are still needed to counter drift and tunnel vision.
• The earlier design, where equal agents coordinated through a shared file and locks, was abandoned.

Handoffs

Planners create tasks. Workers pick those tasks up and push changes. They do not hand work to other workers. The judge decides whether the cycle continues. The post does not describe a message format between planner and worker.

Results

What worked

The post reports a from-scratch browser in close to a week, with over 1 million lines across 1,000 files; a Solid-to-React migration in the Cursor codebase over three weeks (+266K/-193K) that was passing CI and still needed review; a Rust video path the post says was 25x faster and was merged; and other running experiments (Java LSP, a Windows 7 emulator, and Excel) with the commit and line counts given in the post. These figures are the vendor's report.

What broke

The shared-file lock design failed. Agents held locks too long, forgot to release them, or updated the file without the lock, and the lack of hierarchy led agents to avoid hard tasks. The post also says agents occasionally run for far too long. An integrator role for quality control and conflicts was removed because it became a bottleneck.

Lessons

The post says role separation and a fresh start each cycle addressed the coordination failures, that removing the integrator role was an improvement, and that prompts mattered more than the harness alone. It also says fresh starts are still required because drift remains.

Prompts

To be added.

Source

• Wilson Lin, Cursor research blog, January 14, 2026. Vendor source.

  https://cursor.com/blog/scaling-agents

Notes