Which Agent Components Must Stay on Your Infrastructure?
Compare Cursor Self-Hosted Machines with Hermes in Docker against a component-by-component placement policy. Require observed evidence for the agent loop, inference, tools, credential handling, and retained records before approving either deployment.
Classify the control plane before comparing workers
Cursor Self-Hosted Machines puts tool execution on hardware you control while Cursor retains the agent loop and inference. That split can fit a repository checkout rule and still fail a rule covering prompts or tool results. Start with one synthetic task that reads a fixture containing a unique identifier and returns only its line count. Record where the checkout lives, where the read executes, where the model request is processed, and where the tool response is retained. Use worker telemetry to identify connections and use configuration or service-side evidence to identify remote processing and retention. A local checkout and an observed HTTPS connection establish different facts. If a mandatory placement is undocumented or unobserved, carry it as a release blocker in the comparison.
Choose where Cloud Agents run | Cursor DocsInclude persistent state in the ownership check
Running Hermes itself in Docker lets you place the agent process on your host. Its documented data mount at /opt/data holds mutable state, including sessions, logs, and configuration. Inventory the host directory, backup destination, model endpoint, and any log collector as separate destinations. For a staging drill, run a synthetic task with a recognizable run ID, replace the container while keeping the data mount, and retrieve the prior session by that ID. Investigate a missing session as a persistence failure and a session found in an unapproved backup as a placement failure. When testing credential ownership, record where the credential is injected and which service validates it. Use a disposable test account, log only a credential version identifier, and check whether switching versions requires recreating the worker.
Hermes Docker Setup | Hermes AgentDecide which process must stay local
Cursor offers Cursor-hosted Cloud Agents and Self-Hosted Machines. The second option moves tool calls to hardware you control while Cursor still runs the agent loop in its cloud. If your rule covers the repository checkout and where commands run, a self-hosted worker is a candidate. If it covers where the agent decides what to do next, rule that option out.
Before provisioning, write down the allowed location for the checkout, tool process, agent loop, inference, credentials, and audit records. Assign someone to verify each location with configuration, records, or a controlled test. Treat any required boundary without evidence as an open acceptance item.
Choose where Cloud Agents run | Cursor DocsTrace one synthetic tool call
Cursor’s self-hosted worker opens an outbound HTTPS connection to Cursor, which sends tool calls over that connection. That gives you two places to verify a run: the worker’s execution records and the cloud service’s handling of the request. An outbound connection record establishes a destination. To determine what the service receives or retains, verify the documented handling of tool inputs, outputs, and artifacts for your configuration.
For an illustrative trial, put a file containing the synthetic string QA-SAMPLE-17 in a disposable checkout. Ask the agent to read it and create a second file containing only the character count. Check for the expected value of 12, a matching worker event, and the expected network destination. Fail the trial if execution cannot be attributed to the designated worker or if the data flow remains unresolved under your policy.
Integrations | Cursor DocsExercise the credential rotation path
Hermes Agent documents a Docker mode in which the agent process runs inside the container and its user data, including API keys and sessions, lives in a host-mounted directory at /opt/data. For a boundary that requires a locally operated agent loop, test that mode and verify the configured inference destination separately. Use a dedicated test credential and give the mounted directory a named owner with defined access and backup rules.
Recreate the container against the same test directory and verify that expected session data survives. Then revoke the test key and exercise a request that should require it. Verify the old key fails, install its replacement, and use provider-side usage records to confirm the next successful request used the intended account. Investigate any unexpected success after revocation before granting production access.
Hermes Docker Setup | Hermes AgentProve you can investigate a failed run
Daytona Bring Your Own Compute places sandbox workloads on your infrastructure under Daytona’s control plane. Compare it with a self-operated Hermes deployment if your priority is control over sandbox placement and you can accept external management. For either option, decide which records your team must be able to collect and retain when an agent run fails.
Run a successful task and then repeat it against a deliberately missing fixture. Try to reconstruct both runs using a run ID, worker identity, action, target, and outcome. Check that the record contains enough detail for diagnosis without exposing credentials. If the worker log shows the error but your records cannot connect it to the initiating run, add a correlation ID and repeat the test before relying on that setup for incident review.
Bring Your Own Compute (BYOC) | Daytona