Cursor Gives Cloud Agents a Way to Come Back to the Job

Cursor’s cloud agents can now wait for activity outside the chat, return to a task when it arrives, and continue working toward a longer-lived objective. The changes push the product beyond a single uninterrupted coding run and toward a workflow that can react to development feedback.

The trigger moves into the work queue

Subscriptions let cloud agents monitor pull requests, Slack threads and scheduled tasks, then wake when activity occurs. That makes an event in a connected workflow—not just a new prompt—the trigger for another pass at a task.

For pull requests an agent creates, the subscription happens automatically. The agent can then work through CI failures and bot comments until the pull request is complete, keeping the pull request as the object of work after its initial creation.

Goals and modes split the task from the playbook

The new /goal command gives an agent an objective it keeps pursuing until completion. A Custom Mode does something different: it pins a selected skill in the chat, keeping that skill active while the agent works.

The two controls can be combined so an agent follows a chosen playbook while pursuing a durable target. Goals can also work with /loop for recurring check-ins; one supplied example is fixing flaky tests and making CI green in a new chat.

Parallel work gets separate computers

Subagents can now run on their own virtual machines. Each machine receives an isolated project copy and clean context, allowing one worker to test a parent agent’s changes in a fresh environment while others handle independent fixes without colliding.

That separation also supports bug-testing swarms, with multiple subagents examining an app in their own environments. It addresses a practical constraint of parallel coding work: workers need room to test or change a project without sharing the same working context.

New direction waits for the next tool call

Users can send a steering message while an agent is working without cutting off its current action. The follow-up waits for the next tool call, creating a defined point at which the agent can receive a correction or new instruction.

A follow-up can be sent with Send now or by pressing Enter twice. Together, the release adds event triggers, durable goals, isolated workers and queued steering to extended cloud-agent work without requiring each step to begin as a fresh chat request.