LangSmith turns agent traces into fine-tunes

LangChain launched LangSmith Fine-Tuning and its smithtune CLI in public beta; in one code-review test, supervised fine-tuning lifted F1 from 48.9% to 53.7% while cutting model calls 29.8%. The launch results also show 29.4% fewer tool requests, using curated agent trajectories as training data. Teams with repeated workflows should treat production traces as a fine-tuning candidate, not proof of generalization; Snorkel’s data-acceptance contract is the adjacent archive lesson.