mediumAI Engineering

How do you design human-in-the-loop approval for agent actions?

Q: How do you design human-in-the-loop approval for agent actions?

**TL;DR.** Classify actions by **reversibility and blast radius**: auto-execute safe reads, batch-notify low-risk writes, require **explicit approval** for irreversible or outward-facing actions. Approvals must show **enough context to judge**, support **edit-then-approve**, and remember **scoped grants** — or users rubber-stamp everything and the gate is theatre. auto-execute reads, safe queries notify reversible writes, dige

705 views

Understand the problem

Approval gates that catch the risky 2% without drowning users in confirmation fatigue.

human-in-the-loopapprovalssafetyagents

Attempt it yourself

Sketch your approach before reading the solution — that's what interviews test.

Stuck? AI Nudge Available

Get a conceptual hint to guide your logic without spoiling the final implementation.

Study the solution

The solution is waiting

Give it an honest attempt first — then compare your thinking with the full walkthrough.

Read the code

Approval request with preview and scope

const decision = await approvals.request({
  action: "send_email",
  reason: agent.rationale,                     // why the agent wants this
  preview: { to: draft.to, subject: draft.subject, body: draft.body },
  editable: ["subject", "body"],               // human can fix, then approve
  scopeOffer: { kind: "session", pattern: "send_email:to=*@ourcompany.com" },
  approver: routeByAuthority(draft.to),        // account owner, not "anyone"
});

if (decision.status === "approved") await mailer.send(decision.finalPayload);
else agent.observe("Email rejected: " + decision.note);   // feedback, not silence

Join the discussion

Discussion (0)

No responses yet. Be the first to share what you think.

How do temperature, top-p and other sampling parameters affect output?

What chunking strategies exist and how does chunk size affect RAG quality?