AI agents

Autonomy, approvals & undo

Agents are fast, and sometimes wrong. Basemind gives you three safety nets for work done inside a project: an autonomy level, an activity log with approvals, and one-click undo.

Autonomy levels

Choose the level on the project’s Overview tab (workspace admins only).

Level Reads Writes
Safe Allowed Rejected. The project is read-only for agents.
Assisted (default) Allowed Stored as pending. Nothing changes until a person approves. The agent receives pending_confirmation and a preview.
Autonomous Allowed Applied immediately, logged, and can be undone.

An MCP endpoint with Full access behaves as Autonomous on Assisted projects. Safe projects stay read-only for every endpoint on purpose.

The Agent activity tab

Every project-scoped tool call is logged with:

  • the tool, the source (MCP or in-app assistant) and the user it acted for;
  • the arguments and a human-readable preview of the change;
  • the result or error (results of read tools are never stored);
  • the status: pending, executed, failed, rejected or undone;
  • who approved, rejected or undid it, and when.

Filter by status or tool to find what you need.

Approving and rejecting

  1. Open Agent activity and look at the pending action’s preview.
  2. Click Approve to run it, or Reject to drop it.

Details worth knowing:

  • Approval runs the change as the user the agent acted for, with their permissions checked again.
  • If the project changed since the preview was made, the preview is recomputed and the action stays pending, so you only ever approve what you actually saw.
  • Approval is refused while the project is in Safe mode.
  • You decide on actions of your own agents. Deciding on other people’s agents’ actions requires workspace admin rights.
  • The agent can call get_agent_action to see whether you approved.

Undo

Every executed write runs in its own undo session. Click Undo on an executed action to revert everything that tool call did (tables, fields, rows, memories, reports) in one step, all or nothing. Undo is available as long as Basemind keeps the undo history for those actions.

  • Start new projects in Assisted mode and read a few previews to learn how your agent works.
  • Switch to Autonomous for trusted, repetitive work, and keep an eye on the activity log.
  • Use Safe for projects with sensitive or finished data that agents should only analyse.