What happened
Cursor shipped long-running agents as a research preview — autonomous coding agents that plan first, then execute multi-step tasks without waiting for human input at each turn.
Unlike Cursor’s existing inline and chat-based AI features, long-running agents operate over longer time horizons. They break down complex tasks into a plan, then work through that plan autonomously. The result: Cursor can now handle larger, more difficult coding work that previously required constant back-and-forth between developer and AI.
This is a research preview, not a stable release. Cursor is rolling it out to gather feedback before wider availability.
NOTE
“Research preview” in Cursor’s terminology means the feature is functional but expected to change significantly based on user feedback. Previous research previews (like background agents) have taken weeks to months before reaching general availability.
Why it matters
Cursor already had 800K+ monthly active users in early 2025. Adding autonomous long-horizon agents puts it in direct competition with tools like Devin, Factory, and Codegen that have positioned themselves as “AI software engineers” capable of working independently.
The key difference: Cursor’s agents live inside the IDE developers already use. No context-switching to a separate web app. No uploading repos to a third-party service. The agent runs where the code lives.
Planning before execution is the critical detail here. Earlier agentic coding tools often failed on complex tasks because they’d start writing code immediately, hit a wall, and spiral. A plan-first approach lets the agent reason about architecture, dependencies, and sequencing before touching any files — the same workflow a senior developer follows when tackling unfamiliar work.
TIP
Start with well-scoped tasks that have clear success criteria — “add pagination to the /users endpoint with 20 items per page” rather than “refactor the backend.” Autonomous agents perform best when the goal is concrete and verifiable.
Who should care
Developers already using Cursor who regularly spend time on multi-file refactors, feature implementations spanning several components, or boilerplate-heavy tasks. If you’ve been using Cursor’s chat or Composer mode and wished it could just keep going without you approving every step, this is that.
Team leads evaluating whether to add a dedicated AI coding agent (Devin, Codegen) alongside their existing IDE setup should test this first. If Cursor’s built-in agents handle 80% of the use case, a separate tool may not be worth the integration overhead.
Practical takeaway
Check Cursor’s settings for the research preview toggle — long-running agents may require opt-in. When you try it, give the agent a task you’d normally break into 3-5 manual steps and watch how it plans. Review the plan before it executes. That review step is where you catch misunderstandings early, before the agent writes 400 lines in the wrong direction.