Microsoft Research used the DELEGATE-52 benchmark to evaluate how AI models maintain document fidelity during extended, multi-step workflows across 52 professional domains. The study found that frontier models degrade significantly during long chains of delegated work, where they must iteratively edit and reconstruct documents or code.
The results showed an average document fidelity loss of approximately 25% after 20 delegated iterations. Across all tested models and domains, the overall degradation figure reached roughly 50%. Catastrophic corruption, where fidelity scores drop to 80% or lower, occurred in over 80% of the evaluated model-domain combinations.
The testing included Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4. Gemini 3.1 Pro was considered ready for only 11 of the 52 professional domains. However, Python workflows proved to be a rare exception, maintaining stability with less than 1% degradation during extended runs.
The study revealed that providing agents with additional capabilities, such as code execution and file read/write access, actually worsened reliability. The use of agentic tools increased document fidelity degradation by an average of 6%.
Microsoft noted that current verification and orchestration methods have not solved the problem. The research highlights a gap between short-horizon benchmarks used for marketing and actual performance in long-horizon tasks, where a model scoring 95% on a five-step task might drop to 50% on a twenty-step version.
Industry efforts to address these reliability issues in extended agent operations have led to the emergence of frameworks like AgentRx and monitoring tools such as SentinelBench.
