Core Objects

  • Run - A single execution. Identified by runId (UUID). See lifecycle below. Response shape: RunDto.

  • Intent - Declarative outcome. Keep it user-facing and testable (what, not how).

  • Integrations - Tools the agent may call. Use tenant-scoped customerIntegrationIds when creating runs. integrationId is the global key.

  • Context - Hidden execution parameters: timezone, filters, IDs, safety limits, feature flags. Do not expose to end users.

  • UserID - Identity whose tokens/permissions apply (act-as-user). Treat as sensitive. Revoke tokens at the integration.

  • Approval Strategy - NONE, ON_START, or ON_TOOL_CALL for human-in-the-loop control.

  • Plan - Either text or JSON, included in CreateRunResponse. May be displayed for ON_START approvals.

  • Events (SSE) - Real-time stream for status, step start/complete, approvals, errors, and final result.

  • Result - Structured object (if produced) included in RunDto.result.

  • Feedback - Thumbs-up/down and comments attached to a run for evaluation loops.

PreviousMental ModelNextIntegrations

Last updated