Mental model
An agent budget is a multidimensional control envelope. Termination occurs when the goal is verified, progress stalls, a risk boundary is reached, or any resource budget is exhausted.
Learning outcomes
- Explain the mechanism and ownership boundaries behind Agent Execution Budgets and Termination.
- Compare the main design alternatives and their operational trade-offs.
- Diagnose common failures and select evidence for a production decision.
Theory
Track wall time, model tokens, tool calls, retries, monetary cost, depth, and consequence class independently. Stop policies should emit a typed reason and enough state for a human or later run to continue safely.
Trade-offs
Tight budgets control cost and runaway behavior but can interrupt valid complex work. Adaptive budgets improve completion but require trustworthy progress and risk signals.
Failure modes and misconceptions
Resetting counters on retry; using token count as the only limit; allowing the model to extend its own budget; failing to detect repeated states; and reporting timeout as success.
Decision scenario
An agent keeps alternating between two search queries while claiming progress. Define deterministic loop and budget conditions that terminate the run.
Which budget dimensions are needed beyond a maximum number of model turns?
Primary sources
anthropic-effective-agentssre-book