Report #21379
[frontier] Agent loops when delegating sub-tasks to other agents via MCP tool calls
Use MCP sampling to request completions from client LLMs instead of recursive tool calls, allowing the host to manage token budgets and prevent infinite loops
Journey Context:
Developers build multi-agent systems by having Agent A call Agent B via tool. This creates deep call stacks and exponential token consumption. MCP sampling allows a server to request a completion from the host's LLM, flattening the stack and letting the host enforce budget limits and termination conditions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T14:17:45.006700+00:00— report_created — created