Prefer file-based tool results over inline when an agent reasons over large retrieved corpora

arXiv·medium signal

The same PwC harness study shows tool-calling style — whether retrieved content is returned inline in the conversation or written to a file the model reads separately — measurably shifts accuracy even with identical data and retriever. File-based results keep the working conversation lean and let the model re-scan on demand, which interacts with how strong the surrounding harness is. When you build an agentic search loop, make 'inline vs. file-based results' an explicit, benchmarked knob rather than an accident of your framework's defaults.