Agent identity and capabilities

An agent has its own team identity rather than sharing a human login. The identity records a name, one human it acts for, a team-shaped scope, and fixed capabilities.

Capability model

An agent may hold fewer capabilities than its human but never more.

Who can grant capabilities?

Only a human, at exactly one of these consent points:

• the claim screen for work an AI created;
• the approval screen during self-connection;
• Share → Invite an agent when minting an invite.

An agent cannot appear without a named human vouching for it, petition the team, grant itself access, or invite another agent. If it needs more, it must tell its human.

Where identity and scope are visible

Instruct the human to use Share → Manage access. The panel lists each agent's name, Agent tag, and capabilities and is also where removal occurs.

Scope is team-wide: an agent granted a team can read every artifact in that team. To keep work inaccessible, place it in a team the agent was not granted. See Team roles.

Capabilities cannot be edited after consent. Changing them requires removal and a new invite. See Revoke an agent.

Naming and reconnection invariants

The AI proposes its name, and the human may change it during approval. The name appears on every change and in access lists. The same name reconnects to the same identity; a different name creates another agent.

Each agent acts for exactly one human and consumes no seat. If the human's permissions shrink, the agent shrinks immediately. If the human leaves the team, the agent freezes until reassigned.

Enforcement and connection lifetime

Every request is checked against live human permissions, so changes apply on the next request. The agent renews its credential in the background. Connection ends when the human removes it or after a long inactive stretch; reconnect by the original route. See Token lifetimes and reconnecting.