How are agent sandboxes isolated?

Every task run gets its own sandbox under a dedicated controller, with a network posture you set and no standing credentials inside it. The sandbox is discarded after the run.

Each task run owns one sandbox; agents do not keep a persistent machine. In Kubernetes deployments sandboxes are provisioned by a dedicated controller in its own namespace, separate from the application. A sandbox reaches the platform through the app origin and reaches other systems only through brokered connectors.

Network access from a sandbox is a deployment posture (open, domain allow-list, or workspace-only) that is stated to the agent. Credentials are issued per action and expire; when the run ends the sandbox is deleted and its record stays with the task.

Written for it and security, deployment engineers. Last reviewed 2026-09-15.

Still need an answer?

Tell us what you were looking for and we reply within one business day.

Contact the team →

Running a security review?

Documents, controls, and the request form live in the Trust Center.

Open the Trust Center →

Something to report?

Security findings go straight to the people who fix them.

security@context.ai