Error messages built for human developers are quietly sabotaging the AI agents that rely on them.
A paper posted to arXiv (arXiv:2609.35381, revised September 30, 2026) audited 150 widely used Model Context Protocol servers and found 949 of 3,001 error messages tell the caller what to do next. Half of those instructions assume the caller can see something an agent cannot, like its own terminal or a login page. On credential failures, 62 of 67 suggested fixes involve running a terminal command, editing a config file, or opening a web page. On rate limits, 20 of 30 just say to wait, without naming the call to retry. The researchers tested five OpenAI models, including GPT-5.5 and GPT-6 Astra, on Berkeley Function Calling Leaderboard tasks, and the agents dutifully tried to follow instructions they had no way to execute.
The twist: more capable models suffered more, not less. A bare terminal-command instruction let only 45% of expired-credential tasks recover, and the resulting performance hit grew from 18 points on GPT-5.5 to 69 points on GPT-6 Astra. GitHub's stock "wait before retrying" message left just 6% of rate-limited tasks recoverable. Two fixes worked well: naming an actual tool to call instead of a human action pushed recovery to 84-88%, and simply stripping the instruction with a one-sentence prompt before the model read it recovered 82%.
That second fix costs nothing and needs no cooperation from the API provider, which says more about how fixable this problem is than about how capable these agents supposedly are.