It would be a better demo if every agent acted end to end with no human in the path. We deliberately don't build that, and the reason isn't caution for its own sake.
Two categories of action
Actions divide cleanly on one question: how expensive is being wrong. Enriching an indicator, opening a case, or closing a high-confidence false positive with a recorded reason are all cheap to be wrong about. Purging mail from four hundred mailboxes, disabling an executive's account, or pushing a change to production aren't.
| Cheap to be wrong | Expensive to be wrong |
|---|---|
| Enrich an indicator | Purge mail from a mailbox |
| Open or group a case | Disable an account |
| Close a high-confidence false positive | Push a change to production |
| Place an account under monitoring | Block a sender tenant-wide |
The cost of the gate
The gate costs time, and we should be honest about that. An incident waiting on approval at 03:00 waits until someone is awake. We reduce that cost in two ways instead of pretending it doesn't exist.
- The gate covers a set of actions, not each action individually, so one approval releases the whole response.
- Remediation classes can be pre-approved once a team trusts them, which moves work from gated to automatic without changing the audit guarantee.
The result is that autonomy grows with trust rather than being asserted on day one. Every team we have worked with has moved classes from gated to automatic within a few months. None of them would have adopted the tool if it started there.