Sandra Arato

Shipping together in an AI native team

While building an AI agent for marketing workflows at Leonardo.Ai, we discovered that the traditional design-to-engineering handoff was failing. A static mock could describe a happy path, but not how an agent would group assets, select tools, recover from errors, or change behaviour when its underlying model changed. Engineering could implement the interface correctly while still shipping the wrong behaviour.

We began replacing handoffs with shared, executable evidence. Designers built working prototypes inside the product repository and used preview deployments for review. Engineering constrained generative UI through typed tool contracts and a closed registry of design-system components, rather than allowing an LLM to emit arbitrary JSX. We added full-path tracing, tiered evaluators and repeatable datasets so design, research and engineering could inspect the same behaviour. Finally, we packaged that evaluation knowledge as repository-local AI skills, designed to let contributors test ideas without depending on an evaluation specialist.

The results included image and video generation, shipping eight days after we aligned on a reusable prototype path. One model comparison found tool-error rates ranging from 2.0% to 8.1%, including 45 retries caused by a single runaway trace - evidence that a cheaper model substitution was also a product-quality decision. The wider team ultimately reduced agent cost per user turn from 10 cents to 1 cent.

This talk explains why broader authorship needs sharper contracts, observability and accountability - not fewer boundaries.