New in Agentic Eval: Trajectory Capture for Humans and Agents

We've been building tools to capture how humans and AI agents actually complete tasks, step by step, with enough detail to train or evaluate against. This demo, from product team members Sara, Ben, and Nora, walks through two of them.

Human trajectory capture (desktop app)

Records a participant's screen, audio narration, and keyboard/mouse input while they work through a task. In the demo, that's booking a hotel in London against a specific brief. Sessions can be paused and resumed, so longer or multi-day tasks are supported. Playback syncs the video with a caption transcript and input events.

Agentic trajectory capture (web)

Captures an AI agent's run on the web in the same underlying data format, then breaks it into semantic stages so a reviewer can see where the agent was productive versus stuck in recovery. In the demo, an agent asked to plan a Berlin trip spends more than half of its 167 steps in recovery, including a long stretch fighting a broken date picker. The tool lets a reviewer pinpoint exactly where and how it went wrong.

Both products output the same schema, so human and agent trajectories can be compared directly.

The video runs under 10 minutes.

If you would like to find out more, get in touch.