Agent timeouts and budgets
Hard limits on agent runs: max steps, max tool calls, max tokens, wall-clock timeouts, and spend caps.
What it is
Hard limits on agent runs: max steps, max tool calls, max tokens, wall-clock timeouts, and spend caps.
Why it matters
Without budgets, agents loop, burn money, or hold locks forever.
How it works (plain)
Set budgets in the orchestrator—not in polite prompt text. On breach: stop, report, escalate to human.
Try it
Add max_steps=8 and a $ cap to a toy agent loop; force a breach.
Myths
- ⚠️ Myth: “Please don’t loop” is a budget.
- ✓ Reality: Enforcement belongs in code.
Sources
- Course 10 orchestration; permissions; Course 07 cost controls
- OWASP LLM Top 10: https://owasp.org/www-project-top-10-for-large-language-model-applications/ ↗