← Go back

Guide // MandateGuard

How does an AI agent kill-switch work? Freezing spend instantly

By the Ingenious AI team · Published 2026-11-14 · Reviewed 2026-11-14

▸ Run the free live demo See pricing

Free covers 1 agent and 1 policy with the full authorize flow, signed receipts and the approval inbox — no card required to start.

Short answer

An agent kill-switch is a state change in the policy engine: freeze the agent and every authorize call returns DENY with the frozen reason, stamped into the receipt chain, until an owner unfreezes it. In MandateGuard the freeze lives beside the agent record — one action in the approval inbox or via the API — and it dominates every other check: even a perfect in-policy request is refused while frozen. Use it the moment spending behaviour looks wrong and unexplained. Freezing is reversible and preserves the full history, so it costs little to pull early and everything to pull late.

What is an agent kill-switch?

The kill-switch is MandateGuard's freeze control on an agent record. Freezing flips the agent's state in the policy engine. From that moment the authorize endpoint refuses every request for that agent, returning DENY with the frozen reason and writing the attempt to the receipt chain. Unfreezing is a separate deliberate owner action. The switch controls spending only; it does not delete the agent, its policy or its history.

Why a kill-switch is different from lowering caps

When an agent misbehaves, the tempting fixes are slow: lower the caps, edit the allowlist, redeploy the agent. Each leaves a window where the agent can still spend under the new rules, and each confuses the post-incident record by changing the mandate mid-event. Freeze is built for the incident moment. One action changes the agent's state so that every authorize call is refused with the frozen reason, regardless of vendor, amount or day budget. The mandate itself is untouched, which keeps the receipts interpretable: the boundary did not move; the agent was stopped. When the cause is understood, unfreeze restores exactly the policy that was in force before.

What freeze does inside the engine

Freeze is evaluated as part of the same policy check as caps and allowlist, and it dominates them: a frozen agent's request is denied even if it would otherwise be a textbook ALLOW. The denial is not silent. It is receipted like every decision, with the attempted vendor and amount and the frozen reason — which means attempts made during an incident remain visible and countable. That visibility matters, because an agent that keeps trying to spend while frozen is telling you something about its control loop, and an agent that stops trying after the first denial is behaving as designed. Both facts belong in the incident record.

When to pull it

Pull it when the pattern is wrong and the explanation is not yet known: repeated purchases under a new task reference, amounts creeping toward the cap, a vendor you do not recognise appearing in approvals, or denial storms suggesting a loop. Pull it when a vendor reports suspicious orders, or when the agent's host environment is itself in question — a possibly compromised agent should not hold live spending authority while you investigate its machine. Do not wait for proof of loss. Freezing is reversible and cheap; the expensive outcome is another hour of unchecked spend. The receipts preserve whatever happened before the freeze, so evidence is not lost by acting early.

What freeze does not do

Freeze stops spending through MandateGuard. It does not stop the agent's other capabilities — if the agent can also act through tools you have not brought under the mandate, those paths remain open and should be inventoried in the incident. Freeze also does not repair the cause: a looping retry, a bad tool update, a steering page. And it does not adjudicate past spends; the chain does that job by showing exactly what was allowed, approved and denied. Keeping these boundaries clear prevents the common post-incident error of unfreezing quickly because the dashboard looks calm, while the underlying fault is still armed.

Unfreezing safely

Before unfreezing, read the receipts from the last normal decision through the freeze denials. Identify the first anomalous request and what changed just before it: a tool update, a new task source, a prompt change, a vendor change. Fix that cause, or consciously accept it with a policy change you make on purpose — a corrected allowlist entry, a cap adjusted to a verified new price. Then unfreeze and watch the first hour of receipts. If the anomaly resumes, freeze again and treat it as a deeper agent fault. This freeze-diagnose-fix-watch loop is the incident playbook in miniature, and it works because every step leaves receipts rather than recollections.

Freeze as part of a wider safety posture

The kill-switch earns its keep alongside the other paid controls: duplicate detection catches repeat commitments, vendor trust shows per-vendor behaviour from your own ledger, and evidence export packages the incident for auditors or insurers. Teams running several agents should also practise the motion — a quarterly drill where a named owner freezes a test agent and the team times how long detection-to-freeze takes. Like a fire drill, the value is not the rehearsal itself but discovering, in calm conditions, that the person on call knows where the inbox is and has working credentials.

Five actions to take now

  1. Know who on your team can freeze, and make sure they can reach the inbox fast
  2. Freeze first and investigate second when behaviour looks wrong
  3. Read the receipts around the trigger before unfreezing
  4. Fix the cause in the agent or the policy — not by raising caps reflexively
  5. After the incident, verify the chain and export the evidence pack

Worked example: frozen in nineteen minutes

At 14:03 a provisioning agent's receipts show its normal cloud vendor requesting the usual 14-dollar stacks. At 14:11 the same agent begins requesting 46-dollar amounts from the same vendor under a new task reference, six times in nine minutes. Each is under the 50-dollar per-task cap, so they are allowed — until the owner, alerted by the approval pattern on a related queue, freezes the agent at 14:22. The next three requests are DENY, stamped agent frozen.

Review shows a tool update had changed the default stack size; total extra spend was 138 dollars before the freeze, fully receipted, and the remaining attempts cost nothing. The owner fixes the tool configuration, unfreezes at 15:05, and the agent resumes at its normal 14-dollar shape. The receipts preserve the whole arc — normal, anomaly, freeze, denials, recovery — as one verifiable story.

Frequently asked questions

Does freezing delete the agent or its receipts?

No. Freeze changes state only. Policy, history and chain remain intact and verifiable.

What happens to a pending approval while frozen?

The agent cannot redeem or spend while frozen; every authorize call is refused. Review pending items as part of the incident.

Who can freeze and unfreeze?

The workspace owner through the inbox or API with workspace credentials. It is deliberately not something the agent can do or undo.

How fast does freeze take effect?

On the next authorize call. There is no propagation delay inside the policy engine.

Is the kill-switch a paid feature?

It is part of the paid depth (Pro and Team) alongside duplicate detection and evidence export. The free plan still bounds damage with caps and approvals.

Keep reading

Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides