Primary sourcearXiv: OpenAI
View sourceEvent ·Official first-hand
MCP 错误信息为人类开发者编写,反而最伤强智能体
研究分析了 150 个常用 MCP server 中的 3,001 条错误信息,其中 949 条指示调用者执行下一步操作,一半步骤依赖 server 无法感知的调用方环境。用五个 OpenAI 模型在 Berkeley Function Calling Leaderboard 任务上测试,发现凭证过期时错误信息里的终端命令导致任务恢复率只剩 45%,且损失从 GPT-5.5 的 18 分扩大到 GPT-6 Astra 的 69 分;GitHub 的限流提示 "Wait before retrying." 让恢复率降到 6%。补救方面,MCP 开发者在错误信息中直接指名 server 工具可将恢复率提升到 84%-88%,智能体开发者用一句提示词在模型读取前删除该步骤也能恢复到 82%。
Current read
这篇论文实测了 MCP 报错文案怎么坑智能体:能力越强的模型损失越大,还给了两个能让恢复率翻倍的改法,做 Agent 或 MCP 开发的都该看看。
11 sources44° AI heatLast updated 2026/9/28 15:14:46
Evidence
11 sourcesRelated source 1@koltregaskes
View sourceRelated source 2IT之家
View sourceRelated source 3kimmonismus
View sourceRelated source 4Bindu Reddy
View sourceRelated source 5@OpenAIDevs
View sourceRelated source 6AIGCLINK
View sourceRelated source 7The Rundown AI
View sourceRelated source 8小互
View sourceRelated source 9OpenAI
View sourceRelated source 10OpenRouter
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.