July 2026
An agent, showing its work
At Codegen, most of the design effort went into the trace - the agent’s running account of what it understood, what it’s doing, and where to step in. That’s where trust is won or lost. This is that pattern, rebuilt as a working interaction on a fictional task.
The run is scripted, not a live model - it demonstrates the interaction, not an agent. The hypothesis it carries: when the trace is legible, people get better at directing and supervising the agent - they can predict its next step, and they catch the wrong one early. Reading the plan step before the edits land is the whole game. That skill, not engagement, is what I’d measure.
agent trace · 8 steps
The failure follows the animation, not the network - so wait on the cart's settled event, not a longer sleep. This is the step to read before letting it edit.
Details: the trace idles complete, so it reads as a figure before it’s touched. A replay streams it back in, and the timers pause while the tab is hidden. With reduced motion set, a replay resolves instantly to the composed state. Durations sit in tabular figures so nothing shifts as steps land.