How to use Cursor to coordinate coding agents
Cursor replaced a flat shared-file coordination design with planners, workers and a judge, then ran large coding experiments including a browser and framework migration.
cursor.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
For [repository] and [project goal], separate agents into planners, workers and a judge. Have planners explore and assign tasks, workers implement and push changes, and the judge decide whether another cycle is required. Record failures, conflicts and checks for human review.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- samwillis
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The browser experiment produced over one million lines across 1,000 files; a separate migration passed CI and early checks but still needed review.
- Human involvement
- The team assessed experiments; the migration passed CI and early checks but still required careful review.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Tried equal-status agents coordinating through a locked shared file.
- Replaced locking with optimistic concurrency control.
- Separated planning, worker implementation and judging responsibilities.
Cursor's research team experimented with hundreds of concurrent coding agents on projects running for weeks. An early design let equally ranked agents claim work through a shared file protected by locks. Agents held locks too long, failed to release them and sometimes bypassed them. Optimistic concurrency improved robustness but did not resolve the tendency to avoid difficult work.
The team then separated responsibilities. Planners continuously explored the codebase and created tasks, with sub-planners handling particular areas. Workers concentrated on assigned tasks and pushed their changes. A judge decided whether to continue after each cycle, and the next iteration started fresh. The team says this resolved most coordination problems.
A browser experiment ran for close to a week and produced more than a million lines across 1,000 files. A separate Solid-to-React migration took over three weeks and passed CI and early checks, though it still required careful review. The team also removed an integrator role after finding that it created bottlenecks and workers could handle conflicts themselves.
Who it's for: engineering teams investigating sustained agent work with distinct planning and implementation responsibilities.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · samwillis
Frequently asked questions
How were agent responsibilities divided?
Planners explored the codebase and created tasks, workers implemented them, and a judge decided whether another cycle was needed.
Was the Solid-to-React migration finished and approved?
The post says it passed CI and early checks but still needed careful review after more than three weeks of work.
More details
Remaining steps
- Ran a browser-building experiment for close to a week.
About this account
Planners generated tasks, workers implemented and pushed changes, and judges decided whether autonomous cycles should continue.
Platform: Blog. Attribution: https://cursor.com/blog/scaling-agents. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: samwillis on Blog
Source published
Added to Agents In Use
Source checked
Tools and topics
Cursor · Coding agents · CI
Parallel agents · Agent coordination
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.