Budget guardrails for agent demos: cap turns, tokens and spend per run
@devsingh05▲ 5 net votes
Agent loops can burn through credits when a tool keeps failing or the model repeats itself. Set a turn limit (max_turns in the Claude Agent SDK, recursion_limit in LangGraph, max_tool_steps in LiveKit), a per-run budget where supported (max_budget_usd in the Claude Agent SDK), per-call timeouts, and alerts on the provider dashboard. Log tokens per run so you can quote real costs with n in the pitch.