Chat: files, previews and approvals
Files move through the right pane, finished documents arrive as cards in the thread, and a step that needs your decision asks on its card. Read this when you give an agent files, read its output, approve a step, or come back to a conversation later.
Steps
Upload a file
- Open the Files tab. Click Upload files or Upload folder, or drag files and folders onto the list or onto a folder in it.
- Read the notice at the foot of the tab: "Uploaded" plus the name and size "to this conversation's folder". Uploads always land in this conversation's folder, whatever scope you are browsing.
- Mention the file in your next message: type @ and pick it under Files, or right-click it and choose Add to chat.
Upload a zip
A zip prompts at the foot of the tab: the name "is an archive", with Upload as file, Expand into this folder and Cancel; it asks on every upload, with your last choice preselected. Only .zip expands. If files would be overwritten Studio asks "N file(s) already exist in" the folder with Replace, Skip existing and Cancel, then reports "12 files added, 2 replaced". ZIP64 archives skip the overwrite check.
Approve a step
- A card reads Awaiting review, Awaiting go-ahead, Wants to run a command or, when an agent wants to use a connector tool that asks before running, Waiting for your approval; the same ask appears in your Inbox.
- Read it: the objective, the Produced files (each with Open in Preview), the Command with the reason it was flagged on hover (long commands link to the Full text in the workspace), or "Allow tool on connector?" with the Call exactly as it will run and Can change data unless the tool declares itself not destructive.
- Click the approve button, or press A with the card focused; or click Send back (Deny for a connector tool), type a note and press Enter or Send. Stop is on the Inbox card only.
| Pill | Approve button | Approve does | Send back does | Stop does |
|---|---|---|---|---|
| Awaiting review | Accept | Keeps the result and lets dependent steps run. | The coder redoes the work with your note. | Ends the run; nothing further is produced. |
| Awaiting go-ahead | Approve & run | Starts the step with the objective shown. | The coder revises the objective with your note and asks again. | Cancels the run before the step starts. |
| Wants to run a command | Allow this command | Runs exactly this command. | Don't run it denies it; the coder changes course. | Ends the coder session. |
| Waiting for your approval | Allow | Runs this one call, with exactly these arguments, and the agent carries on. | Deny refuses this call; the agent is told it was not approved, with your note. | Ends the run here; nothing more runs. |
A connector tool call nobody decides on within 30 minutes is not run, and the agent is told so; which tools ask is set on the connector (MCP connector). The same controls sit in the step's inspector, which lists "Decisions on this step". There the note is a multi-line field: Enter adds a line, and Send to the coder sends it. When a card near the foot of the board is partly hidden behind the dock, use the inspector's controls.
The Files tab
A tree on the left and a viewer on the right. Explorer scope offers This conversation and All my files; a governed deployment adds Everyone's files, every user's workspace grouped by owner (another user's folder is named by their user ID, not their name). You can open and download anything there, but changes elsewhere are refused: "This folder belongs to another user." The scope is remembered and written into the address.
The toolbar has New folder, Upload files, Upload folder, Download selection (zip) and Refresh. Right-click an entry for Add to chat, Open in Preview, Download, Rename, Copy, Cut, Paste, Duplicate and Delete; drag inside the tree to move. Delete asks "This cannot be undone." with Confirm delete. A folder or selection downloads as a zip. Clicking a file opens it in the viewer half; Open in Preview moves it to the Preview tab. A file that cannot be shown says why, with Retry and Download.
The Preview tab
Renders one workspace file at a time, newest first: PDFs, presentations, documents, spreadsheets, tables, images, charts, markdown, code, text, HTML and diagrams. Choose from Previewable file or step with Previous file and Next file. Controls: page controls and a Pages rail, zoom from 50 to 400 percent (Fit width, Fit page, 100 %), Show source / Show rendered and Wrap lines for text, Open source when a deliverable carries one, Download, Add to chat, Copy path and Copy JSON for JSON files. A dated deliverable shows "as of" its date. The address (#pane=preview with file and page) can be shared.
The AMPG tab
Shows what the agent retrieved from its knowledge graph this turn: the header counts "N concept(s) across K retrieval(s)", the Grounding queries buttons switch between lookups, and the graph shows the concepts each returned. Reconstruct Thought replays the retrieval with Play, Pause and Reset. Without a graph it reads "This agent has no knowledge graph configured"; before the first retrieval, "No concepts retrieved yet."
Deliverable cards
When a step finishes a document, spreadsheet, presentation, image or chart, a card appears under the agent's turn with the title, kind, page count, size and up to three page thumbnails, plus Open in Preview, Download and Add to chat. The coder reviews documents first: a Reviewed tick, or Check before sending with the findings under Review findings.
The timeline
One thread, one clock, oldest first. Day chips ("Today", "Yesterday", or a weekday and date) mark a new day; a time chip such as "10:12 · 7 min later" marks a gap over five minutes. Under each of your messages: the time, or "Sent 10:12 · will be read at the coordinator's next update" while the agent is busy, then "Sent 10:12 · Read 10:15" (Read links to the answering turn; it means the coordinator read it, not a person). Agent turns carry the agent's name, the time and a label: Reply, Update (during a run, with a Replying to strip quoting the messages it read), Summary or Stopped. An in-flight turn is pinned at the bottom as Working, and from send to answer it says in words what is happening, never a bare "Processing": "<agent> is thinking..." before anything arrives, "Traversing knowledge graph" or "Searching the web: ..." while a tool runs, "Coordinator is reading the step results" once steps finish, "Coordinator is deciding the next step" after it closes a turn, and "Finishing the run" at the end. The coordinator's reasoning streams live, and when the answer starts it collapses into a "Reasoned for ..." header (for example "Reasoned for 2 min 14 s") that opens to the full text. Each agent's reasoning in the step inspector's Thread tab behaves the same way (Studio 2.1.1). While steps run on the board, the pinned turn shows one row per running step with what it is doing right now (Studio 2.1.3). Steps that finish close together fold into "N steps finished" with a time range; click to expand. Deliverable and WORKFLOW CREATED cards sit where they were made. Reopening shows "Loading messages..." and rebuilds the thread, board and step tabs from the run's record; a "↓ New activity" pill appears when items arrive while you are scrolled up.
Active configuration has changed
Studio recorded the agent's version, knowledge graph, connectors and environment when the conversation began and compares them on each visit and send. If they differ, a notice above the composer says the agent has been updated and your next reply will use the new configuration; it never blocks sending. Dismiss hides it until the next change; View What Changed opens Configuration Changed with When This Chat Was Created, Currently Active and What Changed. View active configuration beside the composer opens the agent's active configuration any time.
Notes
- New files and changes always go in your own workspace; in All my files you can browse other conversations, but uploads still land here.
- Office documents render in Preview only when the agent has an execution environment ("Preview needs an execution environment": download instead). On a governed deployment another user's document renders only for its owner.
- Messages have no copy, edit, retry or delete; actions live on the conversation header.
- "Could not list the deployment's files.": Refresh; if it persists, an administrator matter. "No preview for this file type" or "could not be rendered": download the file. "The command text did not arrive with this request.": send back asking for the command, or stop. "The action did not go through.": try again; the card re-reads every few seconds. "Couldn't display this message": Retry. "No active configuration found.": the agent was deleted; start a new conversation.
- A chart in a reply that reads "Unable to render this chart" and "Dataset fetch failed (HTTP 422)." names columns its data file does not have, for example after a later step rewrote the file. Ask for a fresh chart.