Skip to content

Agents

Agents define saved workers, their allowed tools, and the credential checks that apply when they run.

Agents are saved worker definitions in Agents & Workflows. A definition names what an Agent may do. A Run records what it did.

Ownership

An Agent owns its saved instructions, tool choices, and limits. The definition also holds its name, description, and output shape. Open the Console Agent directory at /workflows/agents to find that definition. The editor at /workflows/agents/[agentId] keeps its accepted version and edit history. The Agent creation contract owns the exact fields and permission needed to create one.

Lifecycle

An Agent definition can change while older Runs remain readable. Save & Run uses the saved Agent version and records which version ran. The editor also shows a fingerprint of the saved configuration. A Run acts as the Agent. The member who starts it is recorded for attribution. An Agent must declare that it is Taskable before a Task can select it.

Boundaries

An Agent credential requirement names the use it needs, without carrying a secret. A grant binds that requirement to a Workspace connection. Grants can expire or be revoked. A Workflow or block may narrow an Agent's saved tool scope and limits, but cannot widen them. Tool approvals decide whether an exact call should proceed. They do not grant the permission to make it.

Failures

An Agent may be saved yet unable to run. AGENT_NOT_RUNNABLE means its saved definition lacks the required policy enrollment. A credential refusal can mean the required grant is missing, expired, revoked, or bound to another connection. An approval request means the call is waiting for a decision. Treat the reported cause as the next decision. Saving the definition alone does not clear it.

Recovery

An Agent with AGENT_NOT_RUNNABLE must be recreated through the Console, as the refusal directs. For a credential refusal, have an authorized member inspect the saved requirement and its grant. Restore the intended connection or binding before attempting the call again. Use Reviews and repair when a decision or paused Run needs attention. Keep the earlier Run available for comparison.