Developer finds prompt rebuilding, not AI inference, is the real agent loop bottleneck

A software developer discovered that prompt reconstruction — not language model inference — was consuming the most time in their AI agent loop. By isolating and timing four separate spans (serialize, tool, rebuild, and model), they identified a quadratic prompt-rebuild function as the primary culprit. The developer built a reproducible Python harness using a stubbed model to isolate the rebuild cost from actual network latency. All timing data was logged to a single CSV file, which revealed rebuild time growing significantly faster than the fixed model delay across successive rounds. The project is shared as a replicable lab experiment rather than a production case study.