Testing United 2025 - Milano
Milan closed my biggest speaking year. The question on the slide, whether AI can replace a feature team, stopped being rhetorical within months. Here is what that room asked, and what the experiment answered.
LLMs
Large language models in delivery — capabilities, limits, and using them without getting burned.
TestFest : 2025
My first test:fest talk in 2015 was about testing versus checking. Ten editions later I was on the same stage talking about critical thinking and AI. It was the same question, wearing new clothes.
Gate design 101: where quality checks belong in the loop
Add a review" is an incomplete instruction — a check has a location, and a correct review in the wrong place is a slower bug. The cost-of-a-miss curve, the four seams, and four rules for placing a gate.
The overengineering tax: when three tiers cost more than one
Because that's a real failure mode, and it's the one nobody warns you about. Everyone warns you about doing too little. This is about doing too much, and calling it rigour.
The Complete AI Orchestration Journey: From Claude Web Chat to Production iOS App
The reference notes I wrote in November 2025, before BowSmith shipped: four phases, the original 17-agent team, the prompt template that cut context drift, and the failures. Lightly updated, with pointers to where the thinking went next.
Reviewer design: catching what the architect tier misses
A model's output is shaped by its assumptions, it brings the identical set of assumptions to the check. So it reads its own work and finds it sound, not because the work is sound, but because author and reviewer share a blind spot in exactly the same place.
Stop prompting, start delegating
There's a mode most people never leave, and it has clear symptoms. You write one giant instruction and cram every requirement into it. When the output is wrong, you re-roll the same message with slightly different wording, like shaking a vending machine.
