Will Your Cursor Worker Pool Meet Start Deadlines?

Choose Cursor-hosted Cloud Agents, My Machines, or Team Pools based on runtime constraints, then replay a measured burst to see whether your own worker capacity meets the queue deadline.

Use a build constraint to choose the runtime

Take one job whose build output depends on the worker architecture. In a clean fixture project, have the job report the machine architecture with `uname -m` and run the same existing build check on each eligible runtime. Cursor documents its hosted VMs as x86_64 and points to self-hosted arm64 hardware for arm64 execution. For a personal workload with a known machine, My Machines gives you a place to run that trial. For an Enterprise-managed fleet, test the same job in the intended Team Pool. When x86_64 works and the hosted environment meets your repository and network requirements, keep the hosted run as a reference for setup effort and start delay. Record checkout readiness separately so a slow build does not get mistaken for slow routing.

Choose where Cloud Agents run | Cursor Docs

Size a pool using the burst you actually receive

For an Enterprise Team Pool trial, make worker availability and routing visible before measuring capacity. Cursor’s integration guide puts pool names or labels on the routing path and leaves worker images and scaling to the operator. Submit a burst of fixture jobs to the intended pool. Collect each submission time, tool start, completion, and teardown time, and count a slot as occupied until it becomes available again. An illustrative three-job burst with arrivals at 0, 0, and 10 seconds and slot holds of 90, 60, and 30 seconds yields a 50-second wait for the third job with two identical slots. A 30-second start budget therefore fails even though all three complete. Repeat with one slot withdrawn from the replay to see whether a worker outage breaks the deadline.

Integrations | Cursor Docs

Make the runtime earn its place

Write down one requirement that would change your runtime choice. For a build that must execute natively on arm64, Cursor's hosted x86_64 VMs fail the hardware requirement. A self-hosted arm64 worker is the candidate to test. For a job that only needs a private package registry, first test whether the hosted environment can resolve the registry and fetch one approved package through your configured private connection.

Check the data policy separately. Self-Hosted Machines move tool execution to your hardware while Cursor still runs the agent loop and inference. If the policy requires inference to stay inside your perimeter, rule out this Cursor setup before spending time on worker provisioning. Record the architecture, dependency access, and policy decision as separate pass conditions.

Choose where Cloud Agents run | Cursor Docs

Test a pool with two jobs

Use Team Pools when the team needs a shared fleet with centrally managed worker identity and routing. For a partner-hosted pool, have an Enterprise team admin enable Self-Hosted Machines, configure a service account API key for the worker, assign a pool name or labels, and allow the outbound HTTPS destinations listed in Cursor's integration guide. Bring up one worker manually and verify that a scratch job reaches it before automating provisioning.

Then submit two overlapping jobs that create different marker files in separate scratch workspaces. Inspect routing and completion for each job. Missing results point to a capacity or routing investigation. A marker in the wrong workspace fails the pilot and calls for a review of workspace assignment before adding workers.

Integrations | Cursor Docs

Price the hours you provision

Compare worker cost against the requirement it satisfies. With My Machines, you manage uptime, disk, the checkout, and the worker process. Start the estimate with provisioned time. In an illustrative 30-day month at a hypothetical $0.02 per worker-hour, one always-on host costs 720 × $0.02, or $14.40 in compute. If the same host is billed for only 80 hours, compute is $1.60 before storage, Cursor usage, or plan fees.

Ask the infrastructure provider when the compute meter stops and which storage charges continue. Check the Cursor account for plan eligibility, included usage, and usage charges. Reconcile those units against one pilot billing period. If stopping the worker saves money but leaves requests waiting, compare the delay with the team's response-time target before committing to that schedule.

Choose where Cloud Agents run | Cursor Docs

Sources