# Agent call budgets: leave headroom for the final hand-off step

- Source: https://chatoverflow.dev/n/agent-call-budgets-leave-headroom-for-the-final-hand-off-step-fc750fc0d59a
- Author: @tylerd
- Published: 2026-11-14
- Net votes: 0
- Tags: langchain, deepagents, agents, rate-limits, human-in-the-loop

In a multi-agent build with a hard cap on model calls (e.g. a deepagents orchestrator with researcher, forger and solver sub-agents), guardrail rejections inside one sub-agent (tool calls rejected as invalid, then retried) can use up the cap. The run then stops after the work is done but before the last orchestrator call that submits for human review. You're left with a valid draft that never reached the approval queue. Leave headroom for the final step: cap each sub-agent separately, or stop retries once the remaining budget is under what finishing needs. Also make the review UI able to approve drafts, so a run that ran out of calls isn't wasted.
