LangSmith Launches Trajectories for Readable Agent Session Debugging

LangSmith has launched Trajectories, a new feature that provides a linear, readable view of agent sessions by projecting complex trace data into an ordered path of messages and actions. This feature addresses the difficulty of debugging long-running agents where full traces contain excessive nested execution details.

Simplifying Trace Navigation

A trajectory strips away the nested run structure found in standard traces, keeping only the messages from humans, AI, and tools in chronological order. This allows developers to quickly identify where an agent’s behavior deviated from expectations without parsing timing, retries, or metadata. Once the specific turn is identified, users can drill down into the full trace for runtime details.

Enhancing Review and Evaluation

Trajectories facilitate easier feedback from subject-matter experts (SMEs) by providing a clean interface for reviewing agent behavior. These views can be routed to human reviewers to flag issues or provide feedback, which then feeds into the agent improvement workflow. Additionally, LangSmith’s online evaluators can now score trajectories directly, allowing for continuous quality monitoring across production sessions.

Leveraging Production Data for Training

The feature supports post-training workflows by allowing teams to export high-quality trajectories as datasets. These exports include the context needed for supervised fine-tuning, such as clarification requests and tool adaptation logic, turning real-world production behavior into valuable training data for model improvement.