AIE CODE: OpenAI, ArtCraft, Cognition, Exa, Anthropic, Amp, et al
One conference discussion recommends mining agent traces for reusable skills and evals before fine-tuning, with nearly 80,000 team sessions as source material.
A speaker described using agent-session traces to create reusable skills and then build evals from them. Their team had **almost 80,000 sessions** available as source material.
For builders, the practical sequence is to inspect traces, write a skill for a recurring task, and evaluate it before considering specialized fine-tuning. Traces can also reveal when agents recreate libraries or produce work that needs stronger review.
A speaker described using agent-session traces to create reusable skills and then build evals from them. Their team had **almost 80,000 sessions** available as source material. For builders, the practical sequence is to inspect traces, write a skill for a recurring task, and evaluate it before considering specialized fine-tuning. Traces can also reveal when agents recreate libraries or produce work that needs stronger review. The transcript combines several sessions and is fragmentary, so it does not provide measured results for this workflow. The session count is one speaker’s team context, not a general benchmark.
This gives a practical source for reusable skills and evaluations: inspect accumulated session traces, turn recurring work into skills, and derive evals from those traces. The nearly 80,000 sessions describe one team’s material, not evidence that the approach improves results generally.