Handing a powerful agent to one engineer is a demo. Handing it to twenty-five and keeping quality consistent is an organisational problem. The tooling was the easy 20%.

Standards before speed

Before scaling adoption we agreed on the boring things: how agents are prompted, where grounding context lives, what a "healed" test must prove before it merges, and which flows agents are not allowed to touch. Shared patterns meant an agent-generated test from one team read like one from any other.

Review the agents, not just the code

We treat agent behaviour as a reviewable artefact. Every generation and heal is traceable, and code review covers the reasoning as much as the diff. That kept trust high — engineers adopt what they can inspect, and quietly route around what they can't.

Scaling AI testing is a leadership exercise dressed up as a tooling one. Get the standards and the review loop right, and the velocity takes care of itself.