Conversations

You can talk to any worker. The chat panel on a station is where you talk to the agent holding it — and the quickest way to get work into a workflow is to ask for it there.

Opening one

On a station with an agent, the worker strip shows the agent's name. Click the agent to open the conversation.

Asking for work

Ask for user stories and they are written straight into that column as cards. Ask the agent to start on a card and it runs there.

This is how most templates begin. The first station's agent — a business analyst on a software workflow, a strategist on a business one — has no triggers. You open it, say write me user stories for refunds with a reason code, and the stories appear on the board. Drag one to the next station and the line starts moving.

Which model answers

The conversation uses the installed agent's model policy, then the station and factory policies, then the organization default. Manager conversations skip station inheritance. Configure a manager from Manager settings → Agent settings → Model; other installed agents have the same policy editor in the marketplace. A running reply keeps its chosen chain, including a successful fallback, for its subsequent steps. See model policies.

Working from files

Use the paperclip to attach CSV, TXT, Markdown or JSON files. Each file can be up to 2 MB; CSV files can contain up to 5,000 data rows. Use UTF-8 text with a header row for CSV. Quoted commas and multiline fields are supported. Invalid headers, inconsistent columns and unreadable files return an error before the agent uses them.

Files appear under Conversation files with their name and CSV row count. Uploading alone does not ask the agent to create work. Send a message explaining what you want:

Read this support-ticket CSV. Classify each row as a bug or feature, suggest a priority, and explain uncertain cases before creating any cards.

When you are ready:

Create the agreed tickets in this station and export a CSV summary.

The agent must have permission to create cards. Created tickets carry the source file and row, bug/feature classification and priority. The reply links to the cards. Retrying a row in the same conversation returns the existing result instead of creating another ticket, including after deleting and re-uploading the same file. A long import may reach the action limit for one message; ask the agent to continue. Previously created cards remain saved.

Agents can produce downloadable CSV, text, Markdown and JSON files. CSV exports escape spreadsheet formulas. Download an upload or generated file from Conversation files.

Files are private to the conversation owner. SSO holds the encrypted bytes and accounts for storage; CodeBook holds the conversation metadata and checks access before reading or downloading. Deleting a conversation removes its files and import receipts, but keeps cards it created. Deleting a single file keeps its import receipts so a re-upload does not duplicate work.

Your SSO organization needs working storage and a tool-capable AI model. An upload can succeed even when the configured model is unavailable; the file remains available for a later message.

Conversation history

The left side lists your conversations with this station, newest first, with New conversation above them.

• New conversation starts an empty one.
• Clicking an old one loads its history and continues it.

A conversation is bound to the station, so the agent answers with the skill it has there — including that station's instructions.

Letting the agent use a chat on cards

A conversation is private, and that includes from the agent's own later work: what you explained to the QA agent in chat is not known to it when it works a card an hour later. You can change that per conversation.

Tick Let agent use this chat on cards at the top of a conversation. Tick it on as many conversations as you like; the History list marks each one on cards. Untick it to stop.

What the tick does:

• When. The agent is told the chat exists, and can read it, when it works a card or answers in a card's thread.
• Which agent. Only the agent the chat was with. A chat with Maya is never given to Alex.
• Where. Only in the factory the conversation belongs to.
• For how long. Until you untick it or delete the conversation; the agent's access ends at its next step. If you leave the factory, your chats stop being offered.
• How it is used. As context. The agent is told to read the relevant chats before deciding something a card leaves open or asking a person, and that a chat never overrides the card, its acceptance criteria or the station's rules.

One thing to know before you tick: the agent may repeat parts of a shared chat in card comments, and everyone on the factory can read those.

Privacy

Conversations belong to the person who started them. This is enforced on the server, not in the interface:

• Another member of the workspace cannot list your conversations.
• Another member cannot read one by guessing its address.

Both are covered by tests that run against the application. A conversation you tick on cards is still listed and read only by you; it is the agent that may use it. If you want something to be shared with people, put it on a card — the discussion is the shared surface, and it is the one the next worker reads.

What it is good for

• Getting cards written. Stories, bug reports, a plan broken into sections — one card each.
• Shaping a story before it exists. Talk it through with the analyst first.
• Understanding an agent. Ask how it would approach something and you learn what its skill actually says, faster than reading it.
• Unblocking. Ask the engineer what it needs, then put the answer on the card where the run can see it.

What it is not

A conversation is not the record of the work. A run's verdict, its evidence and the routing that follows all happen on the card, where the next worker and the rest of the team can read them. Talk in chat; decide on the card.