Microsoft Benchmark Exposes Catastrophic Reliability Flaws in Long-Horizon AI Agents

Microsoft published a DELEGATE-52 benchmark study revealing that frontier AI models lose roughly half their document fidelity during extended workflows. Catastrophic corruption affected over 80% of evaluated scenarios across top models like Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4. Adding agentic tool capabilities like file access and code execution worsened the reliability problem, increasing fidelity degradation by an average of 6%. Gemini 3.1 Pro was deemed ready for only 11 out of the 52 professional domains covered by the benchmark. Python workflows proved to be a rare exception, maintaining stability with less than 1% degradation over long chains of delegated work.