Discussion · Blog · Try it now
We’re excited to announce long-running agents, now available in the Cursor web app as a research preview for all Ultra, Teams, and Enterprise plans.
Why we built this:
Frontier models can write great code, but they hit a ceiling on larger tasks. They forget the big picture, lose track of what they’re doing, or stop at partial completion. A slightly wrong assumption early on can snowball into a completely incorrect solution by the end. We wanted agents that could handle the kind of ambitious, multi-hour work that previously required constant human oversight—so we built a custom harness that addresses these failure modes head-on.
How it works:
- Planning before execution: Long-running agents propose a plan and wait for your approval before jumping into code, reducing wrong turns on complex tasks and keeping work aligned with your intent
- Follow-through on larger tasks: Multiple agents check each other’s work against the plan, so the agent stays on track across hours or even days of autonomous work instead of losing the thread
- Production-ready output: In testing, long-running agents produced substantially larger PRs with merge rates comparable to other agents and went further to find edge cases, fix similar occurrences, and create high-coverage tests
This builds on our research on self-driving codebases, where we developed a custom harness that enables agents to take on more difficult work and see it through to completion.
Check out the announcement blog post to learn more about how the harness works and what testers have been building.
It’s available now for Ultra, Teams, and Enterprise users!
To try it out, open up https://cursor.com/agents and use the model picker to select “Long-running.”
We’d also love for you to share your first impressions here!