LangChain says Trajectories in LangSmith can be tried now for debugging long-running agents, giving subject-matter experts a readable behavior view, scoring sessions with online evals, and turning production sessions into fine-tuning datasets.
A paper proposes SkillGym, which turns human-written skills into 2,756 training environments and fine-tunes on 8,364 successful trajectories. It reports that Qwen3.5-35B-A3B in Claude Code then gains 19.10 points on Terminal-Bench 2.1 and reaches 51.47% on skill-assisted SkillsBench v1.1.
Ankush Gola says smithtune pulls, selects, and processes trajectories from LangSmith, trains on FireworksAI_HQ and baseten, and evaluates the results with LangSmith. The code is in the langchain-ai GitHub repository.
Fireworks says it is a launch training partner for LangChain’s LangSmith Fine-Tuning and smithtune, describing a one-CLI path from LangSmith traces to managed SFT on Fireworks, then evaluation and deployment, without separate training infrastructure.
Trained on Hugging Face Jobs for about $1.50, the model proposes a task tag and probability from a dataset's column names and first row. Daniel van Strien says 17 minutes of fine-tuning raised its first-suggestion match with an owner's tag from 10% to 69%.
