An AI run budget cap sets the maximum approved cost, tokens, calls or compute for one agent run. Usage from child agents and tools should be included so delegation cannot bypass the limit.
A hard cap is a final safety control rather than the only optimization. Earlier warnings, context compaction, output reduction and steering can preserve useful work before termination becomes necessary.
ELI5
A run budget cap is a hard limit on how much money, compute, tokens, or tool use one AI run may consume. It includes delegated work so an agent cannot bypass the limit by starting child agents.
For example, a research run may stop when it reaches an approved cost of ten euros. Earlier warnings can help the agent shorten output or finish the most valuable work before the hard limit ends the run.
What happens when an AI run reaches its budget cap?
The control system should stop further chargeable actions, record the reason and preserve enough state for review or safe continuation.
Should a budget cap include subagent spending?
Yes. All delegated model, tool and compute usage should roll up to the originating run's approved budget.
