MCP-Injector：120B 模型应对 MCP 工具故障并不比 20B 强
做了个给 MCP 工具注入故障的代理，测完发现 120B 和 20B 表现差不多，真正影响结果的是框架，做 Agent 评测的建议加故障注入。
研究团队推出 MCP-Injector，向 MCP 工具调用注入故障的代理工具，故障类型基于 1,774 个开源 MCP 仓库的 57,240 条 issue 和 PR 记录校准。在 1,350 次运行中，gpt-oss-20B 和 120B 配合三种 Agent 框架（自研 plan-and-execute、OpenAI Agents SDK、LangChain）测试，任务完成分从 8.08 跌至 3.23（满分 10），下降 60%。120B 相对 20B 的优势不显著（p = 0.96），故障下完成率甚至略低（91.1% 对 95.6%）。框架之间的差异大于模型大小的差异，论文已被 NeurIPS 2026 接收。
Be careful when integrating MCP tools into production.
This work shows that a 120B model handles realistic MCP tool failures no better than a 20B model.
They built MCP-Injector, a proxy that injects faults into MCP tool calls. The fault mix is calibrated on 57,240 issue and pull-request threads from 1,774 open-source MCP repositories, so it reflects how MCP servers fail in practice.
Across 1,350 runs with gpt-oss-20B and 120B and three agent stacks (a custom plan-and-execute loop, OpenAI Agents SDK, LangChain), task-completion score falls 60%, from 8.08 to 3.23 out of 10.
The larger model retains a smaller share of its fault-free score (p = 0.96), and its completion rate under faults is slightly lower: 91.1% vs. 95.6%. Differences between agent stacks are larger than the difference between model sizes.
The paper is accepted at NeurIPS 2026.
The practical takeaway is that if your MCP agents are only tested against tools that always respond correctly, add fault injection to the eval and put the effort into framework-level error handling.
https://t.co/NJsrIjKL4m
Chat with Paper: https://t.co/kkmLEu4zkC
