Long-Running Cloud Agents1 / 2

2 min lesson

Cloud agent, multi-hour run

Say what this means in practice: "The task has a definition of done the agent can verify without you."

The job: you can decide when long-running capability fits and scope a governed cloud task. Start with this check: The task has a definition of done the agent can verify without you. Keep it small enough to check the result yourself instead of taking the summary on trust.

Set the boundary for long-running Cloud Agents

0:30 · narrated

Read this demo as text

The real control is called Long-Running Agents. It is team-scoped, and the product warns these runs can last many hours and cost hundreds of dollars.
Then launch from cursor dot com slash agents with a repository, measurable done criteria, and a review owner. There is no user-facing Grind button in this build — long-running is a governed Cloud Agent workflow.

Set the Long-Running Agents team boundary, read the cost warning, then launch from Cloud Agents web.

Practice next: Write measurable done criteria and a spend boundary before assigning a long-running cloud task.

Simulated Cursor 3.12 (macOS, light) — beta educational reconstruction, not the real product.

Learn more

Full explanation

Cloud agent, multi-hour run

Long-horizon cloud run — migrate a framework, then check the artifacts

Cloud agent, multi-hour run

SayThe job is a real long-hauler: migrate the app to the next major version of its web framework across a few hundred files. That is too long for a single short run, so it goes to a cloud agent that can keep working for hours.

DoScope it with a definition of done the agent can check itself against: the app builds, the existing test suite passes, and no deprecated framework APIs remain. Write those as the acceptance the agent verifies before it stops.

SeeThe agent has a target it can test against on its own, not a vague "finish the migration."

DoHand it to a cloud agent and let it run in its own VM for as long as the job takes, working through the migration in the background while you do other work.

SeeIt runs autonomously and comes back with a branch, instead of a half-finished local session tying up your machine.

DoReview what it returns: the PR diff plus the artifacts, the build log and the test run that show the definition of done was actually met.

SeeThe proof point is a reviewable PR with artifacts you read before merging, and the job was long-horizon, not a five-minute change a normal run would have finished faster.

Learn more

Optional practice

Practice: Cloud agent, multi-hour run

QYou've applied it. What actually proves the work is done, not just a plausible answer?