The trace→eval→harness flywheel compounds agent quality — every production interaction generates its own training data

The flywheel is explicit: more usage → more traces → more evals → better harness. Every trace contains valuable data to produce a potential eval, and every good eval makes the harness better. A trace where the agent made a mistake is an eval case; a trace where a user corrected the agent is even better. This builds on Traces are the universal substrate for agent learning — all three layers consume the same execution logs by showing traces don’t just feed model/harness/context updates — they also generate the very eval suite that powers autonomous improvement.

This flywheel creates the kind of Proprietary feedback loops create moats that widen with every interaction that competitors cannot replicate without equivalent production usage. The practical implication: invest in trace infrastructure early, before you need it for optimization, because every production interaction is generating training data whether you capture it or not. Teams that dogfood agents and directly share trace-linked feedback build shared knowledge of agent behavior that feeds back into the improvement loop. Nadella prescribes the firm-level version of exactly this machinery: Private evals should measure business outcomes that matter — not external benchmarks — private evals against business outcomes plus RL environments trained on internal traces are how a company turns its own usage into a compounding learning loop.