Draft, open for argument

Guardrails before autonomy

Autonomy is granted by the person accountable for risk. Give them a reason.

The rule

The step that stops babysitting agents is rarely blocked by technology. It is blocked by a security lead who cannot say what an agent can reach or what it did. Isolation, scoped credentials and an audit trail are what make it reasonable to say yes.

Signs you are breaking it

  • Agents run with a developer’s full credentials.
  • Auto-approve is banned, or allowed quietly and unevenly.
  • Nobody can answer what an agent touched last Tuesday.

In Lights Out

Sandboxed, governed agents moves no work by itself. It unlocks auto-approve and cuts weekly incident risk from about 42% to 8%.

Back to all factors