LangSmith has added a new feature called Trajectories. It shows a clear, chronological view of an agent session. This helps engineers and subject-matter experts (SMEs) understand what the agent did.
A session can include many user turns, tool calls, retries, and subagent handoffs. Full traces contain all these details, but they can be hard to read. Trajectories simplify this by showing only the messages and actions in order.
Trajectories work with many models and SDKs, including OpenAI, Claude, Codex, and Cursor. They remove nested run structures, focusing on the main path of the session. Each message appears once, making it easier to follow.
Reviewers can use trajectories to find where an agent made a mistake. For example, if an answer is outdated, they can scan the session to find the turn where the problem started. They can then inspect the full trace for details.
Trajectories also help with scoring and improving agents. They can be routed to annotation queues or datasets. Online evaluators can score them, helping teams identify sessions that need review or retraining.
Teams can save high-quality trajectories for post-training. These include system prompts, user messages, tool calls, and outputs. This data helps improve agent behavior over time.
Why it matters
Trajectories make debugging faster and more accurate, saving time and reducing errors in agent development.
What to do
Start using trajectories to review sessions, and route low-scoring ones for further analysis.
Source: https://www.langchain.com/blog/langsmith-trajectories-tracing
