← Go back

Guide // MandateGuard

How do you freeze a misbehaving AI agent?

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

Freezing is the immediate response when an agent's spending pattern turns wrong and unexplained: one owner action in the inbox or via the API, and every authorize call for that agent returns DENY with the frozen reason until you unfreeze it. Freeze first and investigate second — the action is reversible, preserves the full receipt history, and converts an open-ended incident into a fixed set of receipts to diagnose. The investigation reads the chain backwards from the freeze to the first anomalous decision, fixes the cause in the agent or its inputs, and unfreezes deliberately with the first hour watched.

What happens, step by step, when you freeze an agent?

Freezing is MandateGuard's kill-switch action on an agent: while frozen, the authorize endpoint denies every request for that agent with the frozen reason, receipted like any decision. Owners freeze and unfreeze from the approval inbox or via the API with workspace credentials. Freezing changes the agent's state, not its policy or history, so an unfrozen agent resumes under exactly the mandate it had — the incident remains fully readable in the chain.

Recognise the freeze-worthy pattern

Freeze decisions are pattern decisions. Single odd receipts usually have single odd explanations; freeze-worthy signals are shapes: amounts migrating toward the cap, a dead task reference coming back to life, purchase frequency doubling without a workload change, approval requests clustering at odd hours, or denial storms from a retrying tool. The common thread is that spend behaviour changed while the mandate did not. You do not need to know why before you freeze — the freeze is what buys the time to find out, at a price of zero additional spend. Operators who wait for certainty usually purchase their certainty with the agent's continued spending.

The freeze sequence

The sequence is deliberately short. One: freeze the agent in the inbox or via the API. Two: note the time and the receipt sequence at the freeze — everything after it that is stamped frozen is incident exhaust, everything before it is evidence. Three: tell the agent's operators the agent is frozen, so nobody helpfully unfreezes it mid-investigation or files the denials as a MandateGuard bug. Four: read the chain backwards from the freeze to the last normal pattern and identify the first anomalous decision. Five: find what changed just before that decision — tool update, task source, prompt, vendor or data. That five-step sequence, executed in order, is most of incident response; the tooling makes steps one and four fast because the state change is instant and the history is chained.

Investigate against receipts, not recollections

Agent incidents attract folklore: someone remembers a deploy, someone else remembers a vendor email. The chain disciplines the story. Timestamps order the events; task references name the work; amounts quantify the exposure; denial reasons show where the mandate pushed back before the freeze. While frozen, keep the policy untouched — changing caps or allowlist mid-incident rewrites the comparison baseline and can mask the trigger. If the agent attempts to spend while frozen, those stamped attempts are data too: frequency and persistence tell you whether the driver is a loop, a queue of stale tasks or an external instruction the agent is following.

Fix the cause, then unfreeze deliberately

Causes live in the agent or its inputs far more often than in the mandate: stale task queues, retry logic without idempotency, tool defaults changed by an update, briefs that outlived their campaigns. Fix that layer. Only if the investigation proves the policy itself was wrong — a verified new price, a legitimately new vendor — change the policy, as its own deliberate act, recorded in your runbook. Unfreezing is then a decision with a name and a reason attached, made by the owner, not a sigh of relief. After unfreezing, watch the first hour of decisions actively; a recurrence means the cause was misdiagnosed, and the correct response is to freeze again faster, now with a second data set.

Mistakes that make incidents recur

Three mistakes recur across teams. Unfreezing because the dashboard went quiet — frozen agents are quiet by definition, so quiet proves nothing. Raising caps to clear the backlog the incident created — this rewards the anomaly with a wider mandate. And treating the freeze as an embarrassment to keep from finance — the receipts exist either way, and a promptly frozen agent with a clean incident story is a governance success, not a failure. The organisations that handle agent incidents well share the opposite habits: freeze early, narrate from the chain, fix causes in the agent layer, and let policy changes be slow, separate and deliberate.

Drill it before it is real

A freeze that has never been rehearsed will be slow the first time: the on-call discovers their workspace credentials are stale, or that they only have the agent's key, which cannot freeze anything. Run a quarterly drill on a test agent — freeze it, confirm denials arrive stamped as frozen, unfreeze, verify normal decisions resume. Time the sequence and fix the access gaps it reveals. The drill also calibrates the team's trigger: after practising, operators freeze on patterns with less hesitation, because they have felt how reversible the action is. In agent spending, the expensive incidents are never the ones where someone froze too eagerly.

Five actions to take now

  1. Freeze immediately when the pattern is wrong and unexplained
  2. Preserve the scene: read receipts before changing any policy
  3. Unfreeze only after the cause is fixed or consciously accepted
  4. Watch the first hour of receipts after unfreezing
  5. Write the incident into your runbook while it is fresh

Worked example: the stale campaign that kept buying

A content team's media agent normally buys stock assets at 8 to 30 dollars from two allowlisted vendors. On a Thursday the receipts show a new pattern: 44-dollar requests, just under the 50-dollar cap, from an allowlisted vendor but under task references for a campaign that ended in June. The operator freezes the agent within the hour; four further attempts arrive over the next twenty minutes and are denied as frozen.

Investigation finds the agent was re-running a stale campaign brief from its task queue. The operator purges the stale briefs, confirms the approved asset list, and unfreezes. Friday's receipts return to the 8-to-30-dollar shape. Total incident cost: three anomalous purchases at 44 dollars (132 dollars) that the approval-band review flagged for refund discussions — bounded, evidenced and closed, because the freeze converted an open-ended pattern into a fixed set of receipts.

Frequently asked questions

How quickly can I freeze?

Immediately — the control sits in the inbox and API, and takes effect on the next authorize call. Practise finding it before you need it.

Will freezing break the agent's non-spending work?

Freeze governs spending decisions. Other agent capabilities are outside MandateGuard; inventory separately any tool that can move money outside the mandate.

Can the agent unfreeze itself?

No. Freeze state is owner-controlled with workspace credentials.

What if I freeze the wrong agent?

Unfreeze it. The action is reversible, and the freeze period's denials are receipted — a mistaken freeze costs delay, not data.

Is freeze available on all plans?

The kill-switch is part of the paid depth (Pro and Team). Free-plan agents remain bounded by caps, allowlist and approvals while you respond.

Keep reading

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