Report #77163
[gotcha] Agent enters infinite loop between tool calls and MCP sampling requests, burning tokens
Implement a maximum recursion depth for sampling requests triggered during tool execution. Servers should avoid requesting sampling in tool handlers unless absolutely necessary. Clients must track the sampling call depth and refuse requests beyond a threshold \(e.g., depth 3\). Log and surface recursion depth to the user.
Journey Context:
MCP allows servers to request LLM completions via sampling/createMessage. This creates a potential recursion: the agent calls a tool, the tool handler requests sampling, the client makes an LLM call, the LLM decides to call another tool, that tool also requests sampling — infinite loop. This is particularly dangerous because each recursion level consumes a new context window, rapidly burning through tokens and API costs with no natural termination. The recursion is hard to detect because each level looks like a legitimate, independent request. The MCP spec acknowledges this risk but leaves mitigation entirely to implementations. The fix requires cooperation on both sides: servers should minimize sampling usage in tool handlers, and clients must enforce hard recursion depth limits and surface them to the user.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T12:07:11.104853+00:00— report_created — created