長期実行エージェントの研究プレビューを拡大

Blog / productCursor's long-running agents research preview is now available at cursor.com/agents for all Ultra, Teams, and Enterprise users.

The long-running agent is the result of our research on agents working autonomously on more ambitious projects, including the work we shared last month on how Cursor built a web browser.

During that experiment, we saw frontier models fail in predictable ways on long-horizon tasks. We addressed these limitations by creating a custom harness that enables agents to take on more difficult work and see it through to completion.

We released a version of this harness last week as part of a research preview. The results show that long-running agents produced substantially larger PRs with merge rates comparable to other agents.

Talking with participants in our research preview, we heard that long-running agents successfully completed a range of tasks that were previously out of reach for agents. A few example runs from the research preview include:

Building an all-new chat platform integrated with an existing open-source tool (runtime: 36 hours)

Implementing a mobile app based on an existing web app (runtime: 30 hours)

Refactoring an authentication and RBAC system (runtime: 25 hours)

Successfully completing difficult tasks requires frontier intelligence and the right harness. By working with every frontier model and building a custom harness for each, we are in a unique position to build the best scaffolding that leverages the strengths of different models. We found that there are a couple general principles that help us achieve better performance.

When iterating directly with a model, tight prompt-response loops let you monitor the agent and nudge it back on course when needed. When the agent goes off and works on a larger task autonomously, a slightly wrong assumption can turn into a completely incorrect solution by the end.

Long-running agents in Cursor propose a plan and wait for approval instead of immediately jumping into execution, recognizing that upfront alignment reduces the need for follow-ups.

Frontier models can write great code, but often forget the big picture of their task, lose track of what they're doing, or stop at partial completion.

Long-running agents use a plan and multiple different agents checking each other's work in order to follow through on larger, more complex tasks.

Initial participants in the research preview used long-running agents to implement large features, refactor complex systems, fix challenging bugs, overhaul performance, and create high-coverage tests.

I shipped two architecture overhauls. It's an incredible tool for "I don't know if this is possible but I'm curious to see" type work. I can run five in parallel, for everything from creating Mac window managers to stuffing CEF into Tauri.

Agents commonly ran for more than a day, producing PRs that merged with minimal follow-up work. Users could step away, focus on other work, close their laptop, and come back to working solutions.

I planned for this project to take an entire quarter to accomplish. With Cursor long-running agents, that timeline compressed to just a couple days. And I could do two or three additional projects. I can kick-off a 52-hour task that I don't have to babysit and come back to a big PR with 151k lines of code.

Compared to synchronous agents, long-running agents were more thorough in their approach and wrote code that was more production-ready.

The magical part of the new harness is allowing the same model to make something production-ready. I tested the same bug-fix prompt locally vs. with a long-running agent, both with Codex 5.3. The local agent fixed it fairly quickly, but the long-running one went further to find edge cases, fix similar occurrences, and create high-coverage tests.

#Using long-running agents at Cursor

For the last month, we've been testing the limits of long-running agents internally. We’ve used them for experiments to see how far we could push them, as well as for production work on Cursor itself. Here are a few tasks we gave long-running agents that we have since merged.

We asked an agent to optimize a video renderer whose performance was bottlenecking deployment. It completed a full migration to Rust and implemented custom kernels, reproducing identical visual output by working purely from the original logic.

Policy-driven network access for sandboxed code

We needed JSON-driven network policy controls and a local HTTP proxy for sandboxed processes. The proxy needed to be correct across protocols, enforce policy consistently, and fail safely without allowing blocked traffic. The long-running agent created a ten-thousand line PR that had very few issues when we ran a large test suite against it. Follow-up work consisted mainly of changes we didn't specify in our initial request.

Some tasks break CLI agents the moment they hit sudo, especially tasks related to system administr