Guide // MandateGuard
What is the playbook for a runaway AI agent spending incident?
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
The runaway spending playbook runs in a fixed order: contain first (freeze the agent or suspend the intent on pattern alone — the action is reversible and every later attempt becomes a stamped denial), read the hash-chained receipts backwards to the first anomalous decision, classify the cause (agent fault, tool or vendor change, steering, or genuine policy misfit) and fix at that layer, restore authority deliberately with the first hour watched, and write the runbook entry with the exposure total and detection-to-freeze time. Most recurrences trace to skipping the classification step and raising caps for what was actually an agent fault.
What counts as a runaway incident?
A runaway spending incident is a pattern of agent purchases outside the mandate's intent — by volume, velocity, vendor or shape — regardless of whether individual purchases were authorised. The playbook is the ordered response: contain (freeze or suspend), read the receipt chain to the first anomaly, classify the cause, fix at the causing layer, restore authority deliberately, and record the incident. MandateGuard supplies the containment controls and the evidence; the playbook supplies the order.
Step one: contain before you understand
Every minute spent diagnosing an uncontained runaway is paid for by the runaway. Containment is one action: freeze the agent in MandateGuard (or suspend the intent, for shared mandates), which makes every subsequent authorize call a stamped denial. Contain on pattern, not on proof: amounts migrating to the cap, frequency multiplying, unfamiliar task references, denial storms — any shape that changed while the mandate did not. Freezing is reversible and evidence-preserving, so the cost of a false alarm is an hour of delay; the cost of the opposite error compounds by the minute. Notify the agent's operators in the same breath, so the freeze is not mistaken for an outage and helpfully bypassed.
Step two: read the chain to the first anomaly
With spend stopped, the receipts become the investigation. Work backwards from the freeze: the frozen-stamped attempts (how persistent was the driver?), then the last allowed decisions, back to the first anomalous one — the first purchase whose vendor, amount, timing or reference breaks the prior pattern. That first anomaly is the incident's hinge: everything before it is baseline, everything after is consequence. Note what changed just before it in the world: deployments, tool updates, prompt changes, vendor incidents, new task sources, calendar events (a campaign launch explains many shapes). The chain's timestamps and references usually narrow the hinge to one or two candidate changes within minutes.
Step three: classify before you fix
Four cause classes cover nearly all runaways, and each has a different fix. Agent fault: loops, retries without idempotency, stale task queues — fix the agent. Tool or vendor change: a default flipped, an API degraded, prices restructured — fix the integration or the vendor relationship. Steering: pages or tool outputs pushed the agent toward purchases — fix the agent's trust in that source and consider allowlist consequences. Policy misfit: the mandate genuinely did not describe legitimate new work — the rarest class, and the only one whose fix is a policy change. Misclassification is the main way incidents recur: the team that raises caps for an agent fault has paid for the same incident twice, once in spend and once in future exposure.
Step four: restore authority deliberately
Restoration is a decision with a checklist: cause fixed or consciously accepted; backlog reviewed item by item rather than released wholesale (anomaly-generated work is usually not real work); policy unchanged unless the classification demands it; counterparties informed if the incident touched shared mandates. Then unfreeze (or resume the intent) and watch the first hour of decisions actively, with a named watcher. A recurrence in that hour means misdiagnosis — contain again and escalate the investigation's depth. A clean hour, followed by a normal day, closes the active phase. The receipts of the restoration are part of the record: future reviewers should see the incident end as deliberately as it was contained.
Step five: the record and the reckoning
Write the runbook entry the same week: timeline from the chain, classification, fixes applied, total exposure (allowed anomalous spend, separated into recovered, refunded and written off), and the detection story — how the incident was noticed and how long detection-to-freeze took. That last metric is the playbook's health indicator; drills and alerting on denial volume exist to shrink it. Present the reckoning to whoever owns the agent budget without varnish: a contained incident with a clean chain is evidence the control system works, and hiding it forfeits both the credit and the lessons. If the incident revealed an ungated spending path, that path's closure becomes a tracked action with an owner — the single most valuable output an incident can produce.
Prevention: what the playbook asks of calm weeks
Playbooks are won before incidents. The calm-week obligations are short: caps fitted so a runaway meets a wall early (the overnight example contained at a daily cap rather than at a card limit); a freeze drill quarterly with detection-to-freeze timed; denial-volume awareness in whatever channel the team watches; task references good enough that the first anomaly is identifiable at a glance; and an approval roster that includes an out-of-hours name for fleets that spend overnight. None of this is heavy. All of it is the difference between an incident that becomes a story about controls working and one that becomes a story about a statement nobody can explain.
Five actions to take now
- Freeze the agent (or suspend the intent) before analysing anything
- Read the chain backwards to the first anomalous decision
- Classify the cause: agent fault, tool change, steering, or policy misfit
- Fix at the layer that caused it; change policy only if the policy was wrong
- Unfreeze deliberately, watch the first hour, and write the runbook entry
Worked example: the overnight retry storm, contained at 682 dollars
02:10 — denial receipts begin stacking: the same 22-dollar data batch, denied on the daily cap, repeating every four minutes. 02:30 — the on-call, paged by the pattern, freezes the agent. Attempts continue, stamped frozen, until the agent's scheduler is stopped at 03:05. Morning analysis reads the chain backwards: before the denials, 31 identical allowed batches starting at 23:40, when a data provider began returning empty payloads the agent treated as a reason to re-buy.
Classification: agent fault (retry without backoff or success validation), triggered by a vendor's degraded responses. Fixes: the agent's retry logic gains backoff and a payload check; the provider incident is noted in the vendor's trust record; the daily cap — which had contained the night to 682 dollars of an unbounded appetite — stays unchanged, because the policy was right. Unfreeze at 11:00 with the first hour watched. Runbook entry written the same day, including the detection gap: the page came from denial volume, and a freeze drill is scheduled.
Frequently asked questions
What if there is no freeze available (free plan)?
Caps and the daily limit still bound the damage. Additionally, you can disable the agent's payment path in your own tooling while you respond. The paid kill-switch makes containment one action — a strong upgrade argument after any scare.
Should I raise caps after an incident to clear backlog?
Almost never in the incident window. Backlog created by an anomaly is evidence, not workload. Clear it item by item after the cause is fixed.
When do we involve the vendor?
When receipts show vendor-side triggers (degraded responses, duplicate charges, steering pages). Bring receipt numbers — the conversation becomes factual immediately.
Does the playbook differ for shared intents?
Containment is suspend instead of (or alongside) freeze, and counterparties must be notified. The analysis and classification steps are identical.
How do we practise this?
Quarterly freeze drills on a test agent, plus one tabletop run of this playbook against a past near-miss. Detection-to-freeze time is the metric to improve.
Keep reading
- How does an AI agent kill-switch work? Freezing spend instantly
- How do you freeze a misbehaving AI agent?
- How do daily spending caps stop AI agents leaking money?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides