Workspace tool

The Workspace tool gives an agent a private, persistent set of files it can read, write, and edit as working memory. It's built in and needs no credentials or setup: assign it to a bot and the agent gets its own file space to keep notes, drafts, and intermediate results in, both during a run and across runs.

What it's for

An agent often needs somewhere to put work in progress that's larger than a single message: an article it's drafting across many steps, research notes it's gathering before it writes, a running list it's building up, or intermediate output it will reuse later in the same task. The workspace is that scratchpad. Files the agent writes stay put, so it can come back to them on a later step or a later run instead of holding everything in the conversation.

It's different from the other two places an agent keeps information. Knowledge is your content, the material you give the agent to answer from. Memory is what the agent learns about the people it talks to. The workspace is the agent's own working files: private to that agent, managed by the agent, and meant for work in progress rather than facts about a customer.

What the agent can do

With the tool assigned, the agent can read, write, edit, and list files in its private space. There are no configuration fields and no actions for you to pick. The agent decides when to use it, the same way it decides when to search the web or recall a memory.

Adding it to an agent

The Workspace tool is built in, so there's nothing to set up. Open Tools and you'll find Workspace under Built-in tools (alongside web search and memory), or add it directly from the agent's Tools tab. Assign it to the bot and it's available right away: no key, no configuration.

Safety in autonomous runs

Like any tool, the workspace carries a safety classification. It's write-capable (not read-only), but it's not destructive and doesn't send data externally. The files live in the agent's own private space, so there's no external destination and nothing to exfiltrate. In practice it's never blocked in scheduled or triggered (autonomous) runs and needs no approval by default. Interactive chat is never affected either way.

In this version

The workspace is agent-internal for now. The files are the agent's working memory; there's no operator-facing file browser and no way to download them from the dashboard yet. You grant the capability by assigning the tool, and the agent manages its own files from there.