Governed autonomy only works if mistakes are recoverable. Brainis makes reversibility a first-class property of every action.
Reversibility by design
Every verb in the action registry declares whether it is revertable and how risky it is. Low-risk actions get a short undo window, medium-risk longer, high-risk longest - so the more consequential an action, the more time you have to take it back. Genuinely irreversible verbs are the ones gated behind approval-always in the first place.
Reverting a single action
From the action's entry in the audit log or the run detail, choose revert: Brainis restores the previous state of the affected records. The revert itself is logged - history never rewrites, it appends.
Undoing a whole run
When an agent run produced many actions, undo the run from the Fleet: every revertable action in the run is rolled back together, in reverse order. This is the safety net that makes trying an agent at higher autonomy a bounded experiment instead of a leap.
What undo cannot do
Externally visible effects cannot be unsent - a message delivered to Slack stays delivered. This is exactly why outward-facing verbs carry higher risk classes and stricter defaults than internal data changes.
Tip: Before raising an agent's autonomy, check its verbs in the registry view: if everything it does is revertable, the worst case of a bad week is an undo. That asymmetry is the honest basis for delegation.