Reading an agent's sources

Where your workspace has agents' steps and sources, see what an agent did while it worked and open the places a reply was built from, with the lines it rests on.

Updated 21 October 2026

If your workspace has agents' steps and sources, you can see where an agent's answer came from. There is nothing to turn on: it is on for everyone in the workspace.

What an agent does while it works, and the steps folded above its reply, show in every chat, whatever the workspace (see Agents, ownership, and visibility).

When the reply posts, the places it read settle into one quiet spot beside the agent's name. Point at it and it offers to check the sources; press it to open them. On a wide screen they open to the right of the reply, so the two sit side by side; on a narrow one they open under it, and on a phone they rise as a drawer. Pointing at the spot slowly pulses every passage the sources back, so you can see at a glance what rests on what. Press the spot again, or the close button at the top of the sources, to fold them away. The bolts the turn cost sit at the foot.

Before a reply that rests on what the agent read is sent, Bolter checks its figures and facts against those reads. The panel's first line says how that went: "9 claims checked", "9 claims checked, 2 to look at", or "Fixed before sending" when the first answer got something wrong and the agent wrote it again. A reply that could not be checked says "Not checked".

The panel opens on its Claims. A bar shows the mix at a glance, and each claim says how it stands against what the agent read:

• Quoted: the reply uses the source's own words.
• Reworded: the source says the same thing in other words.
• Differs from its source: the source says something else, such as another figure.
• Not in its sources: nothing the agent read says it.
• From the conversation: it rests on the conversation or general knowledge, not on something read this turn.

When some claims differ or are not in their sources, the list starts on just those ("To look at"); the filters above it show the rest. Each claim shows the site it rests on, with its icon, and the words it quotes. Pointing at a claim lights its words in the reply. Opening it shows the passage it was checked against, formatted as on the page, with a link that opens the page at those words where the browser can, why it stands as it does, and how it was checked: whether a figure was missing from every source or a second model compared the two, and how much of the page that check read. A page past what the check could read says so, so a claim it could not find there is not taken for made up.

While the panel is open, a small chip follows each sourced claim in the reply itself: the site's icon, or the icon of the file, query or connector it was read from, and a dot in the colour of how the claim stands. Press a chip to open that claim in the panel. On a phone the chips stay in the reply once the panel closes, since the panel covers the reply while it is open.

The Sources tab lists each thing the agent read: a page, a search, a connector, its own database. Opening a row shows the lines of it the reply rests on, and pointing at a row lights the words of the reply that source backs.

Some sources are listed without their contents: anything read through a tool that can return private details, such as a connected server's raw reply, shows as "This source's output is not shown". A reply that read nothing has no spot.