Work

Agentic Eval Suite

Task-based evals that benchmark AI context products against a no-context baseline and inform launch and roadmap decisions.

Panel showing the eval loop, from task to agent run to scored result to launch decision

The goal

"The demo looked good" is not a launch bar. We needed objective answers to two questions. Does our context actually make agents better at building with us? And when a developer asks an agent to build something in our category, how do we do?

What shipped

My team and I built task-based evals for Google Maps Platform. We benchmark launches against a no-context baseline and use the delta to inform launch and roadmap decisions. The same tasks let us compare retrieval, skills, and agent integrations against one developer job.

What I learned

Measurement turns developer experience investments into decisions. Once an evaluated change has a baseline and a delta, the team can put the effort where the result moves. The operating method starts with real tasks, not polished demos.

← All work