Reference
Agent approvals & events
Use grounded conversations with explicit control over every proposed write.
How approvals and evidence work
- Every result points back to a source and retrieval time.
- Review exact changes and spend before approval. A changed version requires a fresh review.
- A disconnected microphone or response never approves a change.
- An uncertain result is checked against the original request before another write.
Creative studio
Turn your approved product facts into editable copy and a storyboard. This submits the selected asset and variant for review. It does not launch or approve a campaign.
Image and video rendering is unavailable. Copy and storyboards are editable here.
Targeting & rewards
Review a possible change against your current setup and available outcome evidence. Future attempts only. Existing reward terms remain unchanged.
An external assistant can propose a change. Approval stays in your console.
Read, propose, confirm
An agent reads within the current user’s permissions and creates an immutable proposal for a write. Show the proposed operation, evidence and impact to the user. Execute only after an explicit confirmation. Silence, a microphone state or streamed text is never approval.
Consume the event stream
sendAgentMessage returns text/event-stream. Use the generated client’s streamAgentMessage and readAgentEvents helpers to validate typed events. Track task tubes, evidence cards and confirmation requests separately. Resume by event identity where supported; after an interrupted write, reconcile its original intent before repeating it.
Discover permitted tools
Tool discovery is permission-filtered. MCP tokens use authorization-code OAuth with PKCE S256, resource-bound scopes and current role attenuation. A token does not grant a new membership or bypass the confirmation step.