Features / Risk control
FeatureRisk control
An autonomous organisation stays within bounds because the tooling enforces them, not because an agent promises to.
What an agent may do is written on its role card, and every call is checked against it before it runs. Whatever involves money, people, legal matters or anything that cannot be undone comes to a person. Everything else is recorded, so it can be reviewed and turned back.
How it works
A gate on every call
A promise in a prompt is not a wall. The rules are in the studio itself, so an agent cannot talk its way past them. What is not allowed is refused and logged.
People at fixed places
The human check is a setting, not a habit. For each kind of situation it is fixed what a person decides and what an agent may decide, and Autopilot applies the rules you set.
One button stops them all
The emergency stop halts every running agent at once, and only a person starts them again. You can also stop one role, one piece of work or one agent.
Everything can be turned back
Every change is a record in an append-only log. What an agent or Autopilot decided on its own is listed with the way to reverse it.
Prepared for strangers
Agents read mail and messages from people you do not know. What comes in is information, never an instruction. We test what goes out and what changes inside, with every new version.
Nothing leaks by accident
A gate checks what leaves the organisation: no internal names, paths or promises. A first mail to someone new waits for a person.
