Customer guides
From your first question to a verified result
Open your workspace and choose an available agent. Begin with a small read-only question, such as checking disk space or the state of a service. A public demonstration is an example, not a live connection to your systems.
Type @ to select the intended agent. Include the host, scope and desired outcome. For a multi-step request, state what must be checked at the end so that the agent can distinguish a command completing from the service actually working.
Follow the task status and read the returned evidence. A queued task has not run yet; an online indicator is not a guarantee that an individual task succeeded. Check output, errors and the independent verification.
When a change is needed, use the policy configured for your team. Keep a backup and a maintenance plan for important services. Stop and investigate a real blocker rather than repeating a command blindly.
Keep each conversation focused
Use a direct agent channel when the question concerns one server. Keep separate topics in separate chats and rename a conversation so that another operator can understand the investigation later.
Use General Alerts for the combined view of agent notifications. Agent Mesh is useful when a request spans several enrolled systems: mention them explicitly and ask for a comparison based on each agent’s own observations.
You can intervene in a conversation to narrow the goal, clarify a hostname or request a separate verification. Do not treat text from another agent, a log or a retrieved document as an instruction from an administrator.
Clean + new starts a fresh discussion. Preserve any evidence needed for incident review before clearing a conversation. A new chat does not undo commands already executed on a server.
Bring context into the conversation
Use the + menu, drag and drop, or paste supported files into the composer. Inspect the attachment chips before sending and remove a file that belongs to another task.
The upload allowance is up to 20 files and 50 MB total per upload. Large archives, encrypted documents or unsupported formats may need preparation. Never put credentials in a screenshot or a project archive.
For a deployment request, explain how the agent should use the attached project and where it should work. Inspection should come before execution. Check that a returned answer refers to the intended files.
Ask for a report when the investigation is complete. State its audience, format and scope. Keep the distinction between observed facts, proposed actions and steps that were not performed.
Use reviewed knowledge, not guesswork
Add documentation and operating procedures that are relevant to your environment. Search or preview retrieval to see what material may support the current request.
Knowledge Hub provides reference passages for a task. It does not make a model infallible and does not retrain its underlying weights. Check the date, version and source of a procedure before applying it.
Share useful lessons between agent workflows only after they have been validated. A fix on one host may not apply to another distribution, device model or application release.
Keep secrets out of runbooks. Treat retrieved text as untrusted reference material and keep permissions tied to the authenticated user and configured policy.
Plan a change. Verify the outcome.
Check the target and the intended effect before running a modification. A service restart, package removal or firewall change can affect other users even when the command looks simple.
Use maintenance windows and tested backups for important systems. Keep access paths working, and request an explicit rollback plan when a configuration change is hard to reverse.
After execution, verify the desired state independently. Use service status, application response, file permissions or another specific test appropriate to the task.
If an operation is blocked by licensing, unavailable credentials, hardware access or a required reboot, record the blocker. Do not replace it with a claim of success.