Improvement

See what the agent ran, and ask about any line of its answer

Each step the agent takes now names what it did on the row itself — the data it read, the search terms, how much came back — and highlighting any passage in an answer hands it straight back for a follow-up.

#Agent

The step list used to tell you a tool ran; now it tells you what it ran. Every step reads as an action and its subject — the table a query read and how many rows came back, the terms a search used, the file it opened — so you can scan a whole turn without opening anything.

Select any sentence in an answer and a small bar appears: explain it, show the data behind it, or chart it. Whatever you pick quotes that passage back, so the follow-up is about that claim rather than the whole conversation.

While a turn is running, the status line now carries an elapsed clock alongside the step it's on, so a long investigation reads as progress instead of a wait.

The grid agent answering which ERCOT settlement points are priciest right now. Above the answer, a collapsed trace reads 'Ran 2 tools', with a 'Search catalog' step and a 'Query' step naming ercot.lmp_by_settlement_point and 20 rows.
Every step names what it did — the data it read and how much came back — above the answer it produced.