Research shows that storing tool outputs in files rather than the context window degrades agent performance. Using a section index for large reports caused GPT-5.1 and GPT-5.2 to lose up to 11.3 points in accuracy, while Claude models remained stable.
HOW THIS AFFECTS YOU
●
builderYou should avoid offloading tool outputs to external files if accuracy is critical for your agentic workflows.
●
researcherThis highlights a new evaluation dimension: the cost of tool output retrieval strategies on model reasoning.