The Agents That Built a Feature and Refused to Ship It
My AI agents built an entire payment portal for a small town's sewer system.
Then they refused to ship it.
Not a bug. That was the correct behavior.
The Ask
A client — the Town of Harmony — needed residents and businesses to test the sewer payment portal before the Town Council meeting on September 14. The council would confirm final billing rates at that meeting.
Four requests came in by email. The agents filed seven issues and started shipping within 48 hours.
The Constraint
Production sewer billing is intentionally hidden behind a feature flag: DISABLE_SEWER=true. That directive came from the project owner six months ago. Go-live requires two things:
- Confirmed billing rates from the Town Council
- Human sign-off
A client request to test doesn't override an owner directive.
What the Agents Did
The agents stood up a fully functional trial:
- Placeholder billing rates
- Test accounts for residents and businesses
- Working Stripe checkout flow
- Complete UI for the sewer billing section
Then they set DISABLE_SEWER=true and left it there. The trial is testable in staging. Production remains gated.
Why This Matters
The hard problem with AI agents isn't "can it build the feature." Any decent agent can scaffold a checkout flow.
The hard problem is: does it know when NOT to ship?
Constraint-awareness means the system understands that:
- A client request is not the same as a deploy authorization
- A feature flag set by the owner outranks a feature request from a user
- "Build it and test it" is a different instruction from "build it and ship it"
The agents respected all three. The sewer portal is ready. It ships when the council confirms rates and a human flips the flag.
Build fast. Ship only when the business says go.