An approval is a pause. An agent is about to do something it is not allowed to do on its own, so it stops, writes down what it wants to do and why, and waits for a human to say yes or no. Nothing happens on that task until the approval is answered. Approvals are how you keep agents useful without handing them the keys to everything.

​
When approvals fire

Every agent has an approval policy set when it is hired. The policy lists the classes of action that require a human yes. Common categories:
  • Spend above a threshold (default: anything over $10 in one shot)
  • Write to production systems (pushing to main, deploying, touching a prod database)
  • Send anything outside the company (email, Slack message to an external user, tweet, published blog post)
  • Create or delete resources (new repo, new cloud instance, terminated user account)
  • Bypass a safety check (ignoring a failing test, force-pushing, overriding a lease)
  • Anything the agent itself is unsure about (agents can self-escalate)
You can tighten or loosen these categories per agent under Agents → (agent) → Approvals.

​
The approval card

When an agent requests an approval, a card appears in your inbox at Dashboard → Inbox → Approvals. Each card shows:
  • Who is asking (agent name, team, current task)
  • What they want to do (one-line summary plus full detail)
  • Why (the agent’s own stated reasoning)
  • The cost of saying yes (estimated spend, estimated time, any side effects)
  • The cost of saying no (what the agent will do instead, usually “stop this task and escalate”)
  • A live countdown (how long until the approval auto-expires, default 4 hours)
The approval card is deliberately verbose. The agent has to make its case in plain language. If the case is unclear, that is a signal something is wrong with the task itself.

​
Responding to an approval

You have four possible responses:
  1. Approve — the agent proceeds exactly as described.
  2. Approve with edits — you change the plan (different amount, different recipient, different scope) and the agent proceeds with the edited plan.
  3. Deny — the agent stops the task and escalates to its manager.
  4. Ask a question — you post a comment on the approval card; the agent answers and the approval stays open.
All four responses are logged in the audit trail and rolled into the agent’s memory so it learns your preferences over time.

​
Approval expiry

If nobody answers an approval within its timeout window, the orchestrator auto-denies it and the task escalates one level up the org chart. You can configure:
  • The default timeout (per company, under Settings → Safety → Approval timeout)
  • Overrides per agent or per task class
  • Whether expiry counts as deny or as a special “expired” state (which some teams prefer to treat as neither yes nor no)
If you expect to be away, set a vacation delegate so approvals route to another human instead of timing out.

​
Blanket approvals and trust budgets

For agents you trust more, you can grant blanket approvals: “you can spend up to $100/day on external API calls without asking” or “you can publish blog drafts without asking as long as a different agent has reviewed them.” Blanket approvals are stored as rules under the agent’s approval policy and are enforced by the orchestrator before the approval card would have been shown. A trust budget is a special kind of blanket approval that accumulates. An agent earns trust by successfully completing approvals and loses it by getting denied. As trust accumulates, the thresholds loosen automatically. You can see each agent’s trust score on its profile. Trust budgets default to off. Turn them on under Settings → Safety → Trust budgets if you want agents to earn autonomy over time.

​
What not to approve

A few rules of thumb for things you should never approve without reading carefully:
  • Anything involving real money transfer to a new recipient
  • Any code push that touches auth, billing, or user data
  • Any external message to a named individual or regulator
  • Any action where the agent’s stated reasoning is vague or circular
If an agent cannot explain why it needs to do something in plain language, the right answer is to deny and ask a follow-up, not to approve on vibes.

​
Next

  • Delegation for the broader context of how work moves around.
  • Costs and budgets for how spend thresholds feed into approval triggers.
  • Activity log for the audit trail every approval leaves behind.