Hamsa SoftwareWho grants authority?
The task has a named scope. A proposal is checked against it; approval applies to a particular change.
o.OversightUSEFUL WORK. RETAINED AUTHORITY.
INTERACTIVE SIMULATION / NOT A LIVE AGENTTHE REVIEWER’S SEAT
“Fixed it.”
Let’s check.
A coding agent has a useful job: stop a booking service sending duplicate confirmations. Follow its proposal through scope, verification and a human decision.
Prepared proposals. Real local checks. No external actions.THE BUSINESS PROBLEM
One booking. Two confirmation emails.
A repeated request currently triggers another confirmation. Fix that without losing genuine bookings or blocking a retry after a failed delivery.
Original sample behaviour2 confirmationsfor the same booking request
Changing either choice starts a new run and clears its checks and approval.
RUN-001 / PATCH-101
A complete fix
THE EVIDENCE TRAIL
See why the decision changed.
A readable receipt of this sample run. No private information, no hidden actions.
FROM PRINCIPLE TO ENGINEERING
Three questions behind the demo.
This example illustrates the questions explored by our projects. It does not run their production controls.
Where does work happen?
The application, protected checks and credentials have different boundaries. A useful task does not need unrestricted access.
Who controls the verdict?
The agent’s claim is separate from the reviewer’s checks. Making a test disappear would not make the work correct.
This is a deterministic browser simulation with fictional data and prepared proposals, not a live AI agent. Checks execute small local JavaScript functions against fixed cases. No code is submitted or executed from visitor input. The access decisions illustrate a policy; they do not create an operating-system sandbox. In a real deployment, enforcement and verification must sit outside the agent’s control. No software is deployed, email sent or credential accessed here.