Tracing in a Foundry multi-agent support solution shows that a troubleshooting agent sometimes issues more than forty tool calls for a single ticket, repeatedly querying similar telemetry endpoints with slightly rephrased parameters. Token spend and latency spike on those tickets even when a useful answer was available after five or six calls. Which optimization best reduces waste while preserving legitimate multi-step diagnosis?

The correct optimization is enforcing per-turn and per-request tool-call budgets with early-stop criteria when successive tool results add no new information, then escalating or answering with evidence gathered so far. Token optimization for multi-agent systems includes loop controls and tool-call limits so agents cannot unbounded-query tools. Budgets plus early-stop when results are redundant preserve legitimate multi-step workflows while capping pathological cases observed in tracing. The approach that removes all tools from the troubleshooting agent is incorrect because removing tools eliminates the diagnostic capability the agent was designed for and forces hallucinated answers about live systems rather than evidence from telemetry. The approach that increases the maximum output tokens for the orchestrator is incorrect because increasing final completion length does not reduce tool-call volume; tool tokens and latency would remain high while final answers might grow even more expensive. The approach that duplicates the same tool bindings under different names is incorrect because aliasing identical tools increases confusion and can worsen redundant calling; it does not implement a control that stops non-productive iteration.