Skip to main content
Every agent run leaves a trail of the moves it made, in the order it made them. Open a cell to see what it searched, what it read, and the reasoning behind its answer.
A table cell drawer with a Steps taken panel expanded, listing search and read steps each with a duration, and a Reasoning field below explaining the answer.

A cell's Steps taken panel, open on an agent run.

​
Open a run’s trail

1

Open the cell

Click a cell the agent filled in to open its drawer.
2

Expand Steps taken

Click Steps taken to open the timeline of everything the agent did for that answer.

​
Read the trail

Each step names one move, in the order it happened:
  • Searched the web, with the query shown beneath it
  • Read a page, naming the site once it resolves
  • Scanned a site, mapping it before a crawl
  • Crawled a site across more than one page
  • Used a tool, naming which one and its arguments
  • Finished, once it has enough to answer

​
See why it answered that way

A muted line above some steps is the model’s own reasoning for the move that follows, folded to one line until you click it. Below the whole trail, Reasoning explains why the value in the cell is the answer.

​
Good to know

The agent had enough from what you gave it and skipped research entirely. Its Finished step still shows, so the trail is never empty.
Each retry gets its own attempt heading above its trail, so you can tell which steps belong to which try.
The cell has not been run yet, or its value came from a lookup that never had a trail to show in the first place.
Yes. Each step shows its own duration, and the Steps taken heading totals the whole run beside the step count.

​
Next steps

Train an agent

Rate an answer you disagree with and teach the agent to do better.

An agent gave a wrong answer

Work through a specific bad answer step by step.