Your AI agent ran up your API bill overnight. Here is the first hour, in order.

1. Pause the agent everywhere it can run. Stop the live process or session, disable any scheduler, cron entry, queue worker, or auto-restart that could relaunch it, and check for a second copy running under a different name. Runaway spend that survives your first stop is almost always a second copy you forgot.

2. Revoke or rotate the exposed key first if the pattern is not your agent shape. If the usage shows models you never use, hours your agent never runs, or a volume your task could never explain, treat it as a possible leaked key, not a misbehaving agent. Revoke the key in the provider dashboard now and investigate after. If the pattern does match your agent, pausing the agent plus a temporary hard limit may be enough, but rotation is still the cleaner stop, as the FAQ below explains.

3. Set a temporary hard provider limit. Drop the account or project limit to a level that cannot hurt you while you triage. You can raise it later with a clear head. Right now its only job is to make a restart harmless.

4. Save the raw evidence before you touch anything else. Export or screenshot the usage views exactly as they are, with dates and times visible. Dashboards roll up and age out detail. Your future dispute, and your own diagnosis, both depend on the numbers as they looked this morning.

Usage by API key. Separate your agent key from every other key on the account: your own experiments, teammates, production services, old scripts. The overnight spike should belong overwhelmingly to one key. If it does not, you have a wider problem than one agent.

Usage by model. Check which models the spending key called. A spike concentrated on one expensive model tells a different story than a spike spread across everything you use.

Usage by hour. Plot or list the night hour by hour. You are looking for the start time, the shape, and the end. A sharp start followed by a flat plateau is a runaway. A series of bursts is retries or scheduled restarts. Normal work has pauses. Runaway work usually does not.

The loop. In the graph: a sudden start, then a steady, relentless rate for hours with no gaps. In the work itself: the same tool or call repeated with near-identical inputs and no progress. The agent was awake and busy all night and finished nothing.

The retry storm. In the graph: sharp bursts stacked on top of each other, often growing after each failure. In the work itself: one endpoint or tool failing, and the agent retrying immediately, sometimes with parallel retries, instead of backing off or stopping. Look for the first failure time. The storm starts there.

The context window resent on every call. In the graph: a moderate number of calls, each one unusually expensive, with input cost dominating. In the work itself: the agent attaching a huge history, a full document set, or an ever-growing transcript to every single request, so each call repays for all the context before it. Call count looks innocent. Cost per call gives it away.

The leaked key. In the graph: spend at hours your agent never runs, on models you never call, sometimes from a pattern that starts abruptly and ignores your task entirely. In the work itself: nothing, because it was not your agent doing the work. This is why step one told you to revoke first when the shape is foreign. Diagnosis later, key dead now.

Separate key per agent. One agent, one key, never shared with scripts, teammates, or production. Attribution becomes a lookup instead of an argument.

Hard provider limit per key or project. Set the ceiling at a number that would annoy you, not ruin your month, and make sure hitting it stops calls rather than only sending an email.

Per-call receipts for the work itself. Run agent work through the Zambo MCP endpoint at https://zambo.dev/mcp (153 tools) so every call, rejected calls included, leaves a receipt you can walk the next morning or the next minute.

An alert before the cap. Set a usage alert well below the hard limit, routed somewhere you will actually see it at night. The alert buys you a manual stop. The cap is the backstop when you sleep through the alert. You need both.

A max-iteration and progress rule in the task. The agent stops after a set number of steps or repeats, and it must show new progress to continue. Most runaways die on this rule before money becomes the lesson.

Should I rotate the key even if it was my own agent?

Yes, in most cases. Rotating the key stops every copy of the agent at once, including schedulers and retries you forgot about, and it closes the question of whether the key also leaked. Issue a fresh key per agent when you restart, update the one place that agent reads its key from, and delete the old key instead of leaving it active just in case. The small hassle of rotation is the cheapest certainty you will buy today.

How fast do providers answer a bill shock report?

It varies, and no honest page can promise a timeline. Reports filed the same day, with a clear time window and evidence attached, are easier to review than reports filed weeks later from memory. Expect an automated acknowledgement first and a human review later. Keep your agent stopped or capped while you wait, keep the usage exports you saved, and answer any follow-up with the same facts and figures you filed at the start.

What evidence do I attach to a dispute or credit request?

Attach the packet that lets a stranger reconstruct the night: the usage export grouped by API key, by model, and by hour, the exact start and stop times, what the agent was tasked to do, the cause pattern you found, and any per-call records you have. Receipts from calls that ran through Zambo are especially useful because each one carries the task, the timestamp, and a fingerprint of the saved result, so the reviewer can see what the agent actually did instead of only what it cost.

How do I cap one agent without capping my whole account?

Give that agent its own API key, and where your provider supports it, its own project with its own limit. The cap then binds the agent, not your other tools, your team, or your production traffic. Pair that with a budget for the work the agent runs through Zambo, so both rails have their own ceiling. The rule is simple: one agent, one key, one cap. Shared keys make shared accidents.

Will the provider refund runaway agent spend?

Sometimes, and never as a right. Providers sometimes credit obvious runaway spend when you report it fast and show evidence that the pattern was a malfunction rather than work you wanted. Nothing is owed to you, approval is at the provider discretion, and partial credit is a common outcome. File quickly, file factually, and do not spend the expected refund in your head before it arrives.