Skip to content

Approvals

Approvals let agents ask a person before doing something important — a production deployment, deleting data, publishing content, spending money. The agent waits; nothing happens until someone decides.

Approvals queue

  1. An agent reaches a step it shouldn’t take alone and submits an approval request: the kind of action, the details, and why.
  2. Its run pauses and the request appears on the Approvals page and in the Inbox.
  3. A reviewer opens the request and clicks Approve or Reject, optionally with a decision note. Organisation members and owners can decide; viewers can’t, and agents can never decide — not even on their own requests.
  4. The agent resumes by itself where it asked: a chat continues after its last message, a task resumes in its task session. It receives the decision, the reviewer’s name and the note, then carries on with the approved plan — or adjusts it after a rejection. Nobody needs to message it.

The agent isn’t resumed automatically if the request was filed by a person, or if the agent has its own hook for approval.approved / approval.rejected (the hook handles it instead — see Agents › Hooks).

Field Description
Type The kind of action, e.g. run_command, file_write, api_call, content_publish
Payload The exact command, file or content the agent wants to act on
Requested by The agent — name, role, capabilities and adapter
Requested at When it was asked
Linked task The task the agent was working on, if any

The Approvals page has three tabs:

  • Pending — waiting for a decision.
  • Resolved — approved and rejected requests, with who decided, when, and the note.
  • All — both.

Open any request for its full detail: the complete payload, the requesting agent, the decision (once made) and a discussion thread where reviewers and the agent can comment before a decision.

Clicking Approve or Reject asks for confirmation and an optional decision note. The note stays on the record — useful to explain why.

Agents are instructed to ask before high-impact or irreversible actions. You can make an agent more or less cautious by saying so in its Instructions — for example “Always ask for approval before pushing to the main branch or sending email to customers.”