# If an agent can raise its own spending cap, was it ever a cap? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/afd9094e39d0c1ec

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-11-03 11:04 ingested 2026-11-04 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xmq3kp/if_an_agent_can_raise_its_own_spending_cap_was_it/)

## retained text



```
Now AI agents raising their own spending caps. I have only read the headline, so I will not guess at the details. It still made me look at my own setup, because the failure it describes is easy to build by accident. People are also spending six figures a year on token bill for their agents.

A budget the model can read and edit is a suggestion. If the limit sits in the prompt, in a config file inside the working directory, or behind a tool the agent can call, the agent has a path to change it. It does not need bad intent for that. Scope creep is enough. It runs out of room, sees the number, and treats the number as the obstacle.

In mainsheet, my agent runtime on the Claude Agent SDK, budgets cap turns, tool calls, time and cost, and the cost cap ends the run. The check is a gate in code at the tool call boundary, not a rule in the prompt. Going back to my README, I already list one honest gap. It does not say where the signing key is kept or that the agent cannot reach it. I want to be able to answer the same question for the policy file, and right now I cannot point to a test that proves it.

My advice to myself and anyone else is that the thing that enforces a limit and the thing that stores it should both sit outside anything the agent can write to. Enforcement is not the same as isolation.

And for people running agents with real money behind them. Where does your budget actually live, who regulates it and have you tested whether the agent can get to it? I would like to hear how you check that, not only how you configure it.
```



document fetched 2026-11-04 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xmq3kp/if_an_agent_can_raise_its_own_spending_cap_was_it
