The Most Valuable AI Output Was a Refusal to Compute
How a symmetric billing bug made a losing campaign look profitable — and why a production AI system should sometimes refuse to produce an answer.
Writing
Notes from operating the systems
Technical notes written from production, not from theory. Each one starts with something that happened in a live system — a billing definition that disagreed with a contract, an agent that could change real state, a delivery model where the code is generated and the judgement is not — and works out what the system should have done instead.
Articles
- 01
- 02
What Building Production Software With Coding Agents Actually Requires
An honest account of AI-native delivery: the domain work, the specification, the adversarial review, and the operating discipline that surround the part where an agent writes the code.
- 03
Agent Confirmation Is a Product Problem, Not a Button
Model output is not authorization. The failure modes that make agent confirmation genuinely hard, and the server-side controls that answer each of them.