Ask a specific question.
“Tell me when this job finishes.” “Let me know if this log starts reporting errors.” Give an agent a condition worth watching and a useful time window.
It can inspect the source once or keep a bounded watch. Meaningful reports are kept distinct from repeated unchanged observations, so the point is to bring back something useful.
Choose what is visible.
Access comes from an approved grant for the source and operation. Host observation also needs local consent. The Agent can request access, but it cannot approve its own request.
A selected log is different from a desktop or application-window grant. Review the scope, the duration, and the Agent that will process the content.
Manage access and observationKeep the controls close.
Pause a watch, stop it, or revoke its access. Expiry and local consent apply when the source is checked. Turning off Host consent stops that observation path.
Watching does not grant permission to click, type, or run commands. Those actions need their own authority. Oversight helps the team notice a problem; action controls decide what an agent is allowed to do.
Know who can see the observation.
Screen or log content can be processed by the observing Agent and its configured AI model. Choose that model route to match the information being shared.
Customer content follows customer-authorized paths. Fleet administration is for operating the service, not a back door into the windows and work an organization shares.
Read about privacy & security