Introducing Trajectories in LangSmith. A view of agent behavior that everyone on the team can understand at a glance.

@trylatitude already does this and @manuelnavid and team shipping everyday 🔥

Readable trajectories are hugely useful, but I'd separate two questions: 1. What path did the agent take? 2. Did the task actually succeed? The trajectory explains process. A separate acceptance test should prove outcome. A perfect trace can still document a perfectly executed wrong result.

A trajectory view turns agent debugging from replay archaeology into a team habit.

finally a visualization tool so PMs can watch my agent fail in real time

A trajectory that shows the tool calls is useful — you can finally see what the agent tried. It seems the part that’s still missing on a real write is whether someone approved that version of the plan, under a scoped identity, and whether you can read the record back after commit. If the write never went through, does the trajectory say “refused,” or does it just stop?

Telemetry FTW

Can teams redact sensitive tool outputs before sending trajectories to SME annotation queues?