Software & DevOpsDone with Cursor Jan 14, 2026
Keep coding agents running autonomously for long stretches
Cursor's write-up on scaling long-running autonomous coding.
- What They Did
- Cursor ran hundreds of concurrent agents on one project. Flat self-coordination through a shared file with locks, then optimistic concurrency, failed. They switched to planners that explore the code and create tasks, and workers that each grind on one task and push. A judge agent decided at the end of each cycle whether to continue, and the next cycle started fresh.
- How It Went
- A browser built from scratch ran close to a week and produced over 1 million lines across 1,000 files. A Solid-to-React migration took over three weeks and passed CI but still needs careful review.
- Worth Knowing
- Prompting mattered more than the harness or model. Agents sometimes ran far too long, and periodic fresh starts were still needed to counter drift.
Try it yourself with Cursor
I want a background agent to work on [large task, e.g. migrating all API handlers to the new validation library] in my repo [repo] for an extended stretch. Break it into small checkable steps, work through them one by one, run tests after each, and commit after each step that passes. Stop when all handlers are done or after [number, e.g. 3] consecutive failures, and leave me a progress log.