LangChain Adds a Readable View of AI Agent Sessions to LangSmith

An AI agent can give the wrong answer after many turns, tool calls and handoffs. Finding the misstep means following what it did without getting lost in every execution detail. LangChain has launched Trajectories in LangSmith to show that path as a chronological session record. The detailed trace is still there when a reviewer needs to investigate a particular step.

Read the path, then inspect the step

LangSmith records a unit of agent work as a run. Runs from one operation form a trace, and traces from a multi-turn session are linked into a thread. That structure preserves an execution tree, including nested work, timing, retries, inputs and outputs. It is useful for a close investigation, but a reader trying to follow the agent's decisions must navigate all that detail.

A trajectory gives that thread a different shape. It gathers messages and actions from the main agent and its subagents, removes the nested run structure, and shows each message once in the order it first appeared. A reviewer can follow the conversation and tool activity, then open the underlying trace to check how a suspicious step executed. The readable view drops the nested run structure; the underlying trace remains available for a closer look.

A starting point for a wrong answer

LangChain illustrates the debugging job with a support agent that gives a customer a wrong answer after nine turns and 60 messages. In the company's example, a reviewer spots where the agent reused an old tool result instead of fetching current data. The reviewer can then open that part of the trace to see the exact tool input and output, timing and retries. It is an illustration of the workflow, not a measured claim about time saved.

That split between reading and investigating also changes who can take part. A support lead or subject-matter expert may know whether an agent followed policy without needing to interpret nested runs. Teams can send trajectories to annotation queues so those reviewers can score behavior, flag problems and give feedback. Engineers still have the detailed record for questions the session view cannot settle.

From one session to a review queue

Trajectories can also be scored by LangSmith's online evaluators across production sessions. LangChain says run-level scoring of long conversations can repeatedly include earlier context as each turn accumulates. Keeping each message once gives an evaluator a less repetitive account of the agent's path. Teams can filter low-scoring sessions for closer review rather than rely on a single final answer to find a problem.

A useful reviewed session can go further: teams can save it to a dataset or export it for supervised fine-tuning, which trains a model on examples of desired behavior. Trajectories contain prompts, user messages, responses, tool calls and tool outputs. That means an example can capture how the agent reached an answer, not only the answer itself. Selecting which sessions represent good behavior remains a judgment for the team doing the review.

Where it works

The view works with traces sent from LangChain, LangGraph and Deep Agents, as well as agent SDKs including OpenAI and Claude and coding agents including Codex, Claude Code and Cursor. LangChain says Trajectories is available now on all LangSmith plans in the US. For teams already recording agent work there, the immediate change is a simpler place to begin a review—not a replacement for the execution record when the cause of a failure matters.