Work mode

Work mode runs an Assistant as a long-running agent with its own workspace. Instead of answering one message, the agent works through a task step by step — reading and writing files, running commands, and using the Assistant's tools — until it is done, while you follow along or leave it running.

Key features

• Its own environment — each Work session runs in a dedicated worker that only sees the Assistant's working directory: the rest of the project is not there. Files the agent creates stay in the project's filesystem.
• Files and shell — the agent reads, writes and organizes files in the working directory and runs shell commands there.
• The Assistant's tools — MCP servers, data sources, integrations, web search, charts, e-mails and image generation configured on the Assistant are available, as are its documents and skills, always called through MINEO: the worker holds no credentials of its own. Instead of code execution, the agent runs code in its own shell, on its own files.
• Files it writes, in the conversation — each file the agent writes links to it in the project, and the images it writes (PNG, JPG, SVG…) show under its reply, as in a chat.
• Queued prompts — send new instructions while the agent is working; they wait their turn and start when the current run ends.
• A conversation like any other — a Work session appears in the Threads list with a Work badge next to its model, and in the sidebar with a green dot while its worker is up. You can rename, share and delete it like a chat.

Requirements

• Your plan includes Work mode, and it is enabled for your organization (Organization → AI → Work, see Organizations).
• The Assistant has the Work capability on, with a Working directory where the agent works (Capabilities → Work, see Assistants).
• You can see the working directory. The agent can change files in it only if the Assistant has File write access and you can edit that folder; otherwise it works read-only, and attaching files needs edit access too.

Start a Work session

1. Start a new thread, from the Threads sidebar or the project home, and pick an Assistant with Work on.
2. Turn on Work in the message box, next to Tools.
3. Write what the agent should do, attach any files it needs, and send it.

MINEO opens the session page right away. The worker button in the header spins while the worker starts, and the agent starts on your message as soon as it is up. The session is named after its first exchange, like a chat, and stops after a while with an idle agent (see below).

Attached files are saved in the Assistant's working directory, in attachments/<session>/, and the message tells the agent where they are. Images and PDFs that the model can read reach it with your message, as in a chat; the agent opens any other file with its own file and shell tools, like any other file in its workspace. You need write access to the working directory to attach files. Up to 10 files per message.

Working with a session

The session page works like a chat — the agent's reply appears as it is written — with a header built for a running agent:

• Title — the pencil next to it renames the session. A spinner next to the title means the agent is working on a turn, the same spinner the sidebar shows.
• Context ring — how much of the model's context window is in use and what the session has cost so far, as in a chat. Compact summarizes the older part of the conversation, as in a chat; a long session is also compacted on its own between turns, and the agent carries on from the summary.
• Assistant avatar — hover it to see the Assistant and the model in use.
• Worker button — the state of the worker the session runs on, as in a notebook:

Worker button	Means	Click it to
Spinner	The worker is starting	—
Green dot	The worker is up	Open Logs, Restart session (a new worker on the same conversation; the agent keeps the conversation but loses any in-memory state) or Stop (releases the worker; the conversation is kept)
Grey dot + Resume	The session is stopped	Start a new worker and carry on where the conversation left off. Sending a message does the same.
Red Retry	The worker failed	Start it again

• ⋯ menu — the chat's: New Thread, Rename, Share thread, Edit assistant (when you can edit it) and Delete, which stops the worker and removes the session from your list.

To stop the agent's current turn without stopping the session, use Stop next to the prompt, or Stop in the conversation's ⋯ menu in the sidebar. The Work section of that same menu stops the session, or resumes it once stopped, without opening it. A command the agent is running in its shell is let finish before the turn stops. A prompt the agent has not started yet can be taken back from its bubble. Your earlier messages can be retried, edited or deleted as in a chat, while the agent is not working: the conversation is cut there and the agent carries on from the new version. A turn you stopped shows as interrupted, with Continue to carry on from where it stopped and Retry to start it again.

The message box takes the chat's commands: /compact and /new, which opens a new conversation with the same Assistant. If the Assistant uses accounts or MCP servers you have not connected, a notice above the message box lets you connect them before the agent needs them.

The model is the one selected in the message box; you can change it during the session, and it applies from the agent's next turn. Later messages can bring files too: they go to the same attachments/<session>/ folder, and a file with the same name replaces the previous one.

When a session stops, the conversation ends with a line that says why (inactivity, a turn that ran too long, a worker that ended or failed, or who stopped it), when, how long its worker ran and how many tokens it used since it last started. To carry on, write in the message box: the session resumes on a new worker and the agent starts on your message.

info

A session stops by itself when its agent has been waiting, with nothing new from you, for your plan's inactivity time (5 minutes by default). While the agent is working it is never stopped for inactivity, and when it finishes the count starts again. A single turn that runs longer than your plan's maximum (8 hours by default) is stopped as a safety measure, and the conversation says so. Administrators can lower the maximum turn duration in Organization → AI → Work.

Limits and governance

• Work sessions count toward your organization's AI budget and cost limits like any conversation. When a limit is reached, the message box says so, as in a chat, and takes no prompts until it resets.
• Your plan may limit how many sessions can run at the same time, across the organization and per person, and administrators can lower those limits in Organization → AI → Work. Starting a session over a limit tells you so and starts nothing; stop one of yours to start another. Running sessions, and their limit, are listed in Active sessions under the Work type.
• A session uses a worker like a notebook does: it counts toward your organization's workers and monthly CPU hours, and gets the project's worker size.
• A session runs as the user who started it: only they can open the session and send it prompts. A shared link shows a read-only copy of the conversation.
• Tools that require approval in chat ask for it here too: the agent waits on the call, and the session shows the same Approve / Deny card as a chat. Stop ends the wait.
• When an administrator turns Work mode off, no new sessions can be started and running ones stop taking prompts, but they can still be opened and stopped.

See also

• Threads — everyday conversations with an Assistant.
• Assistants — configure the working directory, model and tools an agent uses.