I wrote on this two years ago
basically every company will own their pipeline of using Trajectories to mine knowledge + data to improve (harness eng + finetune) their agents
example of things you can do:
- find good examples, use them to distill a smaller, cheaper model
- make environments/tasks from traces. see real world input data and create tasks + simulated environments that model that real world. those evals double as training data if you choose to RL
- compile reports for humans. agents can make sense of a lot of noisy data that humans can’t. create HTML reports to understand what happened with your agent in the last week
hit us up if any of that sounds kinda cool, there’s a lot of fun stuff to do when you capture and organize traces across all your teams and agents— Viv (@Vtrivedy10) August 14, 2026 some great new LangSmith docs on traces vs threads vs trajectories (new concept)
observability data is no longer just for observability - its also for memory & learning
having a really clear mental model of this data is incredibly helpful!https://t.co/UgPfRcVwZg— Harrison Chase (@hwchase17) August 14, 2026
