Agent governance

Agents in Atlas request human sign-off before taking consequential actions. An approval is a stored request that is decided once, memory is explicit key and value, and recent agent runs are listed for audit. Owners and administrators work from the Approvals and governance screen, and Settings, then Agent governance, adds Teammates, Approvals, Memory and Activity tabs.

• Approval gate: a request moves once from pending to approved or rejected, and the decision records who decided and when
• Memory kept per person, per agent, and per workspace
• Double-decide guarded: a decided approval cannot be re-decided
• The Activity tab lists recent runs of every agent, with the trigger, the tokens used and any error

• Saving a memory fact with the same scope, key, agent and person replaces its value instead of adding a second fact.
• Approvals survive a service restart because they are DB-backed rows, not in-memory queues.
• Approvals can be listed and decided, and teammates enabled or disabled, only by workspace owners and administrators. A template here starts the agent builder. Apart from the Day planner, whose schedule changes wait for approval, a template is never run directly, and no run is recorded without a model.

1. 1Run an agent from the Agents screen. When the run reaches a step that changes data, it stops and shows the step with Approve and Reject.
2. 2Approve or reject the step on the run page. Approving runs the step and the run continues, and rejecting ends the run as failed.
3. 3Approvals raised by governed templates, such as the Day planner, and approved actions waiting to be replayed are handled on the Approvals and governance screen, which only owners and administrators can open. A step that belongs to a run can also be decided from the Approvals queue: approving it there runs the step and resumes the run, exactly as approving it on the run page does.

What stops an agent doing something consequential without me?

An approval gate. A new agent waits for a person's approval on every tool that changes data, unless its builder chose otherwise. The agent proceeds only on approval.

Can an approval be reversed after somebody decides it?

No. A decided approval cannot be decided again. That is deliberate, so the record of who allowed an action and when cannot be rewritten afterwards.

Does an agent forget everything if the service restarts?

No. Memory and approvals are saved records rather than temporary queues, so both survive a restart. Memory is an explicit key and value, kept per person, per agent, or per workspace, and writing the same key replaces the value.