LangChain Launches LangSmith Fine-Tuning CLI for Agent Trajectory Post-Training

LangChain has introduced LangSmith Fine-Tuning, a new capability centered on the smithtune CLI that allows developers to convert agent trajectories stored in LangSmith into supervised fine-tuning (SFT) datasets. This tool streamlines the post-training workflow by handling dataset creation, training execution, and evaluation within a single interface, eliminating the need for custom data pipelines.

The system leverages agent trajectories—sequences of messages, tool calls, and results—to train specialized models. By capturing the precise context the model saw at every step, including dynamic tool availability, smithtune ensures that training data reflects real-world agent behavior. The CLI supports integration with Fireworks and Baseten for the actual model training process.

In internal tests, fine-tuning base models on curated trajectories yielded significant performance improvements. For instance, fine-tuning Kimi K2 improved its score on an internal benchmark from 90.0 (base) to 96.0, outperforming GPT-5.6 Sol's 85.0. Similarly, fine-tuning Qwen3-8B improved its score from 78.0 to 82.0. These results suggest that domain-specific SFT can enhance model performance on targeted tasks.

The workflow involves pulling trajectories from LangSmith, filtering for high-quality examples, and configuring training parameters via the CLI. Users can adjust hyperparameters and select specific model bases before launching the training job. The tool also includes evaluation features to compare the fine-tuned model against the base model, providing immediate feedback on the effectiveness of the post-training run.