Skip to main content

Shared Workspace Files

An agent is only as useful as the material it can reach. Until now that material lived wherever you kept it, which usually meant a folder on one machine that nobody else could see. Agenta now gives you and your agents one folder in the cloud, and your team works from the same copy.

Getting files in​

You can add a file four ways: paste it from your clipboard, drag it in, click the attach button, or drop a whole folder. When you drop a folder, Agenta uploads its contents one file at a time.

Images, PDFs, spreadsheets, and plain documents all work. Large files upload in chunks, so a slow connection does not sink the whole upload, and a file that is too big gives you a clear message instead of a failed request.

Two folders, and the difference matters​

Files you attach to a conversation belong to that conversation. They are the desk for today's work: the spreadsheet you are asking about, the drafts you are shaping, the results you want to take away.

Each agent also has its own folder that persists across every conversation with it. That is where the reference material goes: the brand guide, the examples your team approved, the notes the agent should follow. Point at those files from the agent's instructions and it starts every session already knowing them, instead of you explaining the same context again.

Both are visible to your team, and the agent reads them from inside its sandbox exactly as you see them.

See Files and knowledge for how session files and agent files differ.

Simpler built-in tools for Pi agents​

Pi agents have always had seven built-in tools: read, bash, edit, write, grep, find, and ls. Until now, you managed them like any other tool, listed as entries in the agent's configuration and in the template.

These tools are now always available. They no longer appear in the agent configuration or the template, so there is one less thing to set up when you build an agent.

Control moves to the permissions panel instead. Every agent has a default policy (reads are allowed out of the box), plus allow, ask, and deny rule lists you can edit. These rules work the same way across harnesses, not only Pi.

If an existing agent's configuration still lists built-in tools from before this change, it keeps working. Agenta just ignores those entries now.

See Permissions for how the default policy and rule lists work.

Better-equipped sandboxes​

Agent sandboxes now ship with more everyday shell tools out of the box: ripgrep, fd, jq, unzip, zip, tree, file, process tools, and a python alias for python3. An agent that reaches for one of these no longer needs you to install it first.

See Customize the agent runtime for the full list of what sandboxes ship with.

Bug fixes​

  • Approving a tool call no longer evicts a warm agent session, so resuming after an approval is fast again.
  • Denying a tool call now reliably ends the turn.
  • A resolved approval no longer reappears after you reload the page.
  • The session Files drawer now reliably shows that session's own files.