Agents, Runs, and Approvals

Agents are user-owned automation identities with optional workspace access, skills, memories, and external MCP servers.

Agent lifecycle

1. List agents with GET /api/v1/agents.

2. Read an agent and its workspace bindings.

3. Start a run with POST /api/v1/agents/{agent_id}/runs.

4. Poll GET /api/v1/agent-runs/{run_id}.

5. Read approvals and approve only high-impact actions after human review.

Agent runs execute within the requesting principal's tenant and permission boundary. An agent cannot use its existence as a reason to broaden access.

Production guidance

Use a dedicated agent for each durable workflow. Keep system prompts versioned, set bounded max steps, and use approval gates for deletion, publishing, sharing, billing, and external side effects.