Your AI agent changed the code, ran a few commands, and declared victory. Can you reconstruct what happened? Keeping agent trajectories gives you a record of the prompts, tool calls, and outputs behind the result. Analysing those records can help you understand failed tasks, find recurring mistakes, improve how your team uses agents, and investigate suspected misuse. If you only start collecting after something goes wrong, the evidence may already be gone.
This talk will explore what you can learn from stored trajectories and why they’re worth keeping even before you know which questions you’ll need to ask. I’ll also share lessons from building Quesma Shipper, our open-source trajectory collector: dealing with scattered session data, changing formats, and transcripts full of secrets. We’ll discuss how to preserve useful records without creating an unnecessary store of sensitive data, and where trajectories alone fall short as evidence.
Seasoned software engineer and architect with 15 years of industry experience and a strong background in systems design and operations. Currently a member of the founding team at Quesma, building tools that help engineering teams understand how AI coding agents work, what they cost, and the value they deliver. Previously helped build Elastic’s Cloud offering and delivered crowd testing and mobile app quality tooling at scale at uTest/Applause.
