AI agents can ask. Only a person can say yes.
PACT holds the high-risk actions your AI agents request, like granting admin access. A real person sees the actual action, not the agent's story, and approves or denies it with Circle ID. Everything else runs as usual.
The new risk
The agent isn't hacked. It's convinced.
A capable agent follows instructions. A convinced one follows a convincing attacker. The prompt that talks it into assigning admin rights looks a lot like the prompt that asks it to do its job.
Why current controls don't stop it
Scopes and step-up
They approve a window of time, not the specific action. Once the agent has access, it can do anything in scope.
Monitoring and alerts
They tell you after the action has already happened.
Email and chat sign-offs
A reply in email or Slack doesn't actually block the agent. It can go ahead without one.
What the approver sees
The real action. Not the agent's story.
Your phone shows only the action, so you approve exactly what you read, and nothing the agent wrote can sway you.
Operation
The exact action the agent wants to take, written in plain language.
Context
Where the request came from and who or what asked for it.
Agent's reason
What the agent says it's doing. Marked unverified because the agent wrote it.
Four actions that always need a person.
Admin role assignments
Adding Global Administrator or other privileged roles.
Password and sign-in resets
Resetting credentials or authentication methods.
User deletion
Removing a user account and its data.
New app credentials
Issuing secrets, tokens, or API keys.
The questions you're already asking
Do I have to rewrite my agents?
No. PACT sits between your agents and the systems they touch. You add a gate, not a rewrite.
Can an agent go around it?
No. The action runs only after PACT releases it. The agent can't skip the signature.
Will my agents get stuck waiting?
Only on the actions you gate. Everything else runs the same as before.
Let your agents work. Keep the big calls human.
Launching Q1 2027. Join early access to be first in line, and tell us which agent actions you'd want a person to approve.