Keep Your Agents From Burning Through Your Budget
Autonomous agents can run up real costs through retry loops and runaway work. Set budgets, retry limits, and stop conditions before the spend, not after.
Founder, Task Machine
You handed an agent a job and let it run. The work came back done, more or less. Then you checked the bill. The agent had been busy in ways you never asked for: repeating the same expensive step, calling tools you did not expect, and pushing forward long after a person would have stopped to ask a question.
Nothing crashed, and there was no error to point at. The agent kept working, every minute of that work cost money, and by the time anyone noticed, the spend had already happened.
Overspending rarely looks like a failure while it happens. It looks like progress, right up until you see the number.
Why agents overspend without anyone deciding to
A person doing recurring work has a natural sense of "enough." They notice when a task is taking too long, when they are repeating themselves, or when a small job has somehow become an all-day job, and they pause to ask whether it is still worth it.
An agent has no such instinct unless you give it one. Left alone, it treats "keep going until the task is finished" as a literal instruction. If finishing is hard or impossible, it keeps trying, and each retry, extra tool call, and additional pass costs something. The cost is invisible while it happens and obvious only afterward.
Most agent setups have no built-in answer to a simple question: how much is this allowed to cost? Without that answer, spending is a side effect of letting work run unattended, and nobody ever decided on it.
Cost control is a trust feature
It is tempting to file this under finance, something to reconcile at the end of the month. That framing is too small, because unpredictable cost is the main reason people are afraid to let agents run on their own.
If you cannot predict what a job will cost, you cannot leave it alone. You hover, you babysit overnight runs, and you keep agents on short, supervised tasks because the long, valuable ones feel too risky. The fear of a surprise bill caps how much real work you are willing to hand over.
So cost control is really a reliability feature. A job with a known ceiling is a job you can walk away from. The aim is to make spending something you decide on purpose, so that letting agents work alone stops feeling like a gamble.
The common ways agents run up a bill
Most cost surprises come from a short list of repeatable patterns. Naming them is the first step, because each one has a specific control that stops it.
| Failure mode | What it looks like | What causes it | Control that stops it |
|---|---|---|---|
| Retry loop | The same expensive step runs over and over | A step keeps failing and the agent keeps re-attempting it with no limit | Retry limit |
| Runaway loop | Work continues with no natural end | The task has no clear "done," so the agent never stops | Stop condition or verifier gate |
| Over-eager autonomy | The agent keeps going where a person would have paused to ask | No defined point at which uncertainty should hand back to a human | Approval before spend |
| Expensive tool calls | A few costly actions dominate the bill | The agent uses high-cost tools or models more often than expected | Budget cap |
| Silent escalation of spend | A small job grows into a large one without notice | No running total and no threshold that triggers a check-in | Budget cap plus approval before spend |
Read down the list and a pattern appears. The expensive cases are almost always work that should have stopped earlier, to ask, to check, or to give up, but had no mechanism to do so.
Four controls that turn spend into a decision
You do not need a complex system to fix this. You need four explicit limits, set before the work starts.
- Budget cap. A hard ceiling on what a job may cost. When the job reaches it, work pauses instead of pushing past it.
- Retry limit. A maximum number of attempts for any step. After that, the job stops re-running the expensive thing and surfaces the problem.
- Stop condition. A clear definition of "done," so the agent has a finish line instead of working until something external interrupts it.
- Approval before spend. A point where a person sees the cost decision and says yes or no before the money is spent.
The last one matters most and is most often missing. Most tooling tells you what an agent spent, and far fewer tools stop to ask before the spend happens. By the time a notification says "this run cost more than usual," the money is already gone, so the decision has to come first.
The tradeoff
Tight limits have a cost of their own. A budget cap set too low will stop legitimate work that needed one more pass. A strict retry limit can give up on a step that would have succeeded on the next attempt. An aggressive approval gate can interrupt you for spending you would have happily allowed.
You cannot remove all risk without removing autonomy, and an autonomous job will sometimes need to spend more than you guessed.
What you can do is make the limit explicit and the overage a decision. A job that pauses at its budget and asks whether to continue is in a far better state than one that blew past a number nobody set. You can always approve more, but you cannot un-spend money on work you never agreed to.
Where Task Machine fits
Task Machine is built around that model: recurring work done by humans and agents carries its limits with it, so nobody has to remember to check them afterward.
A job can carry a budget and a retry limit as part of how it runs. When it approaches its ceiling, or needs to spend in a way that crosses a threshold, that decision becomes an item in your inbox, a question that reaches a person before the money moves. You approve it, adjust the limit, or stop the job.
When the work is finished, the run history shows what happened: what was attempted, where it retried, and what it cost. That record lets you tune the next budget instead of guessing again. The inbox keeps in-flight decisions in front of a person, and the history keeps the past honest.
The change is small, and it alters how it feels to let agents work. Spending stops happening to you overnight and becomes something you decide, sometimes in advance and sometimes in the moment, but always on purpose.
Let agents work overnight without the surprise
Most people keep agents on a short leash out of fear of what an unattended run might cost, and removing that fear makes longer, more valuable jobs possible.
Budgets, retry limits, stop conditions, and approvals before spend remove it. They turn an open-ended run into a bounded one, and a surprise bill into a decision you got to make.
If you want to let your agents do real work overnight without checking the bill in the morning with one eye closed, start a 7-day trial.