LangChain has added Trajectories to LangSmith, giving teams a chronological view of what an AI agent did across a multi-turn session. The feature combines messages from users, agents, tools and subagents in the order they first appeared, without making reviewers begin with the full nested execution tree.
Detailed traces remain available for exact tool inputs, outputs, timing and retries. The trajectory acts as an index: a reviewer can find the turn where an agent reused stale data or chose the wrong tool, then jump into the underlying trace for technical evidence. It supports data produced by LangChain, LangGraph and Deep Agents, as well as OpenAI, Claude, Codex, Claude Code and Cursor integrations.
The simpler view is also intended for subject-matter experts who understand whether an outcome was correct but do not need runtime metadata. Teams can send trajectories to annotation queues, score them with online evaluators and save useful sessions as datasets. Each message appears once, which can reduce the repeated context sent to evaluators for long conversations.
High-quality trajectories can later become supervised fine-tuning examples because they preserve system prompts, user requests, tool calls and outputs—not just the final answer. Trajectories are available now on all LangSmith plans in the United States; the announcement does not specify availability elsewhere.