A 28 September 2026 preprint by Xiaonan Xu and Wenjing Wu surveyed 150 widely used MCP servers and found that many error messages suggest next steps, such as a terminal command or a web page, that an agent acting only through tools cannot perform. In the study, the more capable GPT-6 Astra lost more recovery than GPT-5.5 when given such a step.
Rewriting the step to name a server tool restored recovery to 84-88%, and a one-sentence prompt reached 82%. The evidence is one preprint on five OpenAI models, and agents with a shell or a human nearby may be spared.