Files and AI Agents

Agents use the same files you do. An agent can read a spreadsheet from the shared drive, work on it, and save the result into a folder, onto a task, or into its reply. It works with the permissions of the person who called it, so it reaches only the files that person can.

Giving an agent a file

There are two ways, and they can be combined.

Attach it to your message. In a channel, a direct message or a brainstorm, attach the files to the message in which you mention the agent. The agent receives those files with your request, and it can see what else has been attached in the same thread.

Name it by path. For a file that is already in the workspace, tell the agent where it is. In the Files area, Copy path in a file's or folder's menu gives you the address.

The file is… Its path looks like
In a folder of the Files area /invoices/2026/aug/report.pdf
Attached to a task /tasks/BE-12/spec.pdf
Attached to a project /projects/website/brief.docx

For example: "@analyst compare /invoices/2026/aug/report.pdf with the July report in the same folder and save a summary next to them."

What an agent can read

How much an agent gets out of a file depends on the kind of agent.

Agent Text, Markdown, CSV, JSON PDF, Word, Excel, PowerPoint Images
Harness agent Yes Yes. The files are placed on its computer and it opens them with code Yes
Server-side agent on OpenAI or Anthropic Yes Yes. The files are placed in the provider's sandbox Yes, it sees them directly
Server-side agent on another provider Yes No. It knows the file's name, type and size, but not what is inside No
Your own AI client over MCP Yes If the client can download files: it receives a short-lived link Depends on the client

For work on documents and data, such as joining spreadsheets, reading a long PDF or producing a chart, a harness agent is the right choice. See How Agents Work for the three kinds of agent.

What an agent can do

It can Notes
Browse the folder tree and list the files of a folder, task, project or thread It sees names, types, sizes and dates
Read a file By path or from the files attached to the conversation
Create a text file Markdown, CSV, JSON, plain text or code, in a folder, on a task or on a project
Save a file its work produced A chart, a cleaned table, a converted document. Into its reply, or straight into a folder, a task or a project
Copy a file From one place to another; the original stays where it is

If an agent saves to a folder that does not exist yet, the folder is created.

An agent cannot delete, rename, move or replace a file, and it cannot rename or delete a folder. When it saves a new result under a name that is already taken, a second file appears beside the first. Tidying up is yours to do in the Files area.

The same limits apply to agents as to people: programs are refused, a file can be up to 50 MB, and it counts toward the workspace's total storage. See what you can upload.

What you see

  • A file saved into a folder appears in that folder while you have it open, and the agent's reply links to the folder.
  • A file saved into the reply is shown under the agent's message, like any attachment.
  • A file saved onto a task or project appears on its Files tab.

What an agent adds to the Files area is recorded in the workspace activity feed, the same as for people.

From your own AI client

When you connect your own AI client over MCP, it gets the same abilities as tools:

Tool What it does
folder_list Lists the folders of the Files area
file_list Lists the files of a folder, task, project or thread
file_get Reads a file by path or id. Text arrives directly; other files arrive as a short-lived download link
file_create Creates a text file in a folder, task or project
file_copy Copies a file to another folder, task or project
file_upload_begin, file_upload_commit Upload a file from your machine

You do not call these yourself. Ask in plain language, for example "put this report into /reports/2026", and the client picks the tool.