Unverified50% confidenceFactExact time
When function calling fails, one only needs to check the prompt, the model output, and the API response; whereas when an MCP agent fails, one must trace across three layers: the LLM planner, the MCP client, and the remote MCP server.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMCP distinguishes between protocol-level errors (such as connection drops and timeouts) and tool-level errors (such as invalid parameters and insufficient permissions).77% similarVerifiedMCP defines a standard response mechanism for errors occurring during tool calls.76% similarUnverifiedThe MCP tool call success rate is affected by instruction clarity: with short, vague instructions (such as 'list semua pm') the Agent cannot call the tools correctly, while with a complete phrasing (such as 'use IdCloudHost's MCP to check current usage') it succeeds.74% similarUnverifiedMCP协议涵盖服务发现、能力声明、权限管理和错误处理的完整生命周期73% similarUnverifiedWhen throwing exceptions in MCP tools, ToolError should be used instead of ordinary ValueError, because ToolError lets the model see the explicit reason and return value, whereas ordinary exceptions may cause the Agent to see only 'tool execution failed' or even trigger meaningless automatic retries.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1042905API
curl https://kongchang.com/api/v1/knowledge/claims/1042905MCP
get_claim(id=1042905)