Handling approvals¶
When an agent reaches a decision it shouldn't make alone, it creates an approval and waits. Your job is to read it, decide, and let the agent know why.
For the background — when does an approval get created? — see the approvals concept page. This page is the day-to-day mechanics.
Where to find approvals¶
The Approvals page¶
Sidebar → Work → Approvals. Tabs across the top let you filter by status (All, Pending, Approved, Rejected). The Pending count is what you usually want. Click any row to open it.
The Inbox¶
Sidebar → Overview → Inbox. The Inbox is your single "what's waiting on me" view, with tabs for Issues, Approvals, and Failed Runs. The Approvals tab lists everything pending; the sidebar shows a badge for the count.
On the issue itself¶
Every approval links back to the issue the agent was working on. That issue's Timeline shows an "Approval requested" entry — click through to reach the approval from there.
Reading an approval¶
The approval detail page shows:
- Title — the agent's stated reason, at the top with a status badge.
- Description — the agent's fuller explanation, if it gave one.
- Linked Issues — usually one; click through to see exactly what the agent was doing.
- Comments — humans (and the requesting agent) can talk here before you resolve. Use it for "why are you asking?" or "what have you considered?".
- Properties (right side) — status, type, requester, comment count, and when it was created.
Resolving¶
When the approval is pending, the Properties panel shows one short form:
- An Optional context text box.
- An Approve button and a Reject button.
Type your reasoning into the box (optional), then click Approve or Reject. Whatever you type is posted as a comment on every linked issue and is visible to the agent on its next heartbeat. Leaving it blank submits a clean verdict with no note.
Make the note count:
- On Approve, explain why it's OK so the agent learns the pattern. "Approved — this contract template was pre-cleared by legal in 2025-Q4" beats a bare "approved".
- On Reject, say what would have made it OK. "Reject — this needs CFO sign-off, not mine" lets the agent re-route instead of re-asking.
Once resolved, the approval locks. A Delete button appears on resolved approvals (it removes the approval and its comments; the linked issues themselves are preserved).
What happens after you resolve¶
Within the agent's heartbeat interval:
- The requesting agent wakes up.
- It sees the result and your note in its prompt context.
- By convention (built into the role prompts) it will, on Approve, continue and confirm with a comment; on Reject, acknowledge and then re-route or close the issue.
If the agent is paused or its heartbeat is off, nothing happens until you trigger it — Run Heartbeat on the agent page.
Rules of the road¶
- Only humans can resolve. Agents can read approvals through the API but cannot approve or reject them. This is enforced server-side.
- One approval is not blanket permission. Each approval covers the one action the agent named. Approving "send this contract to legal@external.example" does not pre-approve future similar actions. For a standing policy, write it into the agent's system prompt.
- There's no undo. To change your mind, comment on the linked issue (the agent sees it next heartbeat) or wait for the next decision point. An over-eager agent that took an irreversible action right after Approve is a real risk — match how much you trust an approval to how reversible the action is.
Bulk operations¶
There is no "Approve all" button — every approval takes an individual click, by design. You can bulk-clean resolved approvals from the Approvals page once they're closed.
Nudging stale approvals¶
Want a reminder when an approval sits too long? Trimus's reminders come from agents and routines, not a built-in timer on the approval itself. A common pattern is a scheduled routine that checks for old pending approvals and pings a chat channel — set one up under Work → Routines. See the heartbeat page for how scheduled work runs.