Skip to main content

​
Conversations

Each conversation is a persistent thread with the agent, auto-titled from your first message. The agent keeps working while you’re away — long-running turns continue in the background, and if your connection drops mid-answer, the turn keeps going and the stream picks back up. Every conversation gets its own isolated workspace: a sandboxed Linux environment where the agent runs commands, works with files you attach, and queries your organization’s data. Workspaces are per-conversation — nothing leaks between threads.

​
Choosing a model

Fieldwork offers a curated set of frontier models — the latest Claude, GPT, and Gemini models — selectable per conversation from the composer. The default is a strong general-purpose choice; switch to a heavier model for hard reasoning or a faster one for quick lookups. A conversation remembers its model, and the context-usage ring in the composer shows how much of the model’s context window the conversation is using.

​
Approvals

Actions with side effects wait for you:
  • Workspace actions — running commands, writing or editing files in the workspace
  • Knowledge base writes — page creates, edits, and deletions (the approval card shows a diff)
  • Connector actions that change things — creating an invoice, updating a listing
  • Automation changes — creating or modifying any standing scheduled job
Reads are generally exempt: searching the web, reading knowledge base pages, listing your data tables, pulling records from a connector. The approval card always shows the exact command, diff, or payload — what you approve is what runs. Auto-approve is a per-conversation toggle for when you trust the loop (say, a long analysis session in the workspace). It never extends to another conversation.

​
Attachments

Drag files into the composer:
  • Images and PDFs go to the model directly — ask questions about a screenshot or a contract.
  • Data files (CSV, XLSX, and similar) land in the conversation’s workspace, where the agent can parse, transform, and analyze them with real code.
The agent can also produce files — cleaned datasets, reports, charts — and hand them back as downloads.

​
Sharing a conversation

Conversations are private to their owner by default. Share one with teammates as:
  • Viewers — can read the conversation and download its files
  • Editors — can also continue it: send messages, steer, approve
A shared editor’s turns run under their identity and access — sharing a conversation never widens what it can reach. If an editor asks for data they can’t access, the agent can’t get it for them. Archiving and deleting stay owner-only.

​
Steering and stopping

You can queue follow-up messages while the agent works, interrupt a turn to redirect it, or stop it outright. If an approval card is pending, denying it tells the agent to find another way.