Matt Pocock is a leading voice in the TypeScript world. He's known for asking tough questions and demanding clarity from developer tools. And he's skeptical of AI agent frameworks. Most feel like overhyped abstractions to him.
So we brought him into a Mastra workshop. No prep. Just build, break, and react in real time.
Here's a recap of the highlights in case you missed it.
First Contact with Mastra
Matt showed up with questions. "What even is an agent? How is it different from a function?" He called out naming choices, default settings, and edge-case handling.
But when we showed a typed agent with clear input/output and schema validation, his tone shifted. "Okay, this I can work with," he said. Strong TypeScript support didn’t hurt.
Primitives That Pass the Smell Test
We broke down the core primitives: inputs, steps, tools, output checks. Matt saw familiar patterns. Not magic, just structured logic.
"You're not claiming the AI does everything. You're giving me a framework I can reason about. That’s good."
He liked the analogy: agents as interns who need supervision. "Managing AI interns with types. I can live with that."
Evals: Skeptical, Then Sold
Matt didn’t hide it: "Evals are usually fluff. No one shows how to make them useful."
So we showed a working eval setup inside a live workflow. A test agent runs tasks. A judge agent scores output. The score controls retries or path changes.
He tested edge cases. The system held up. "Okay, this is interesting," Matt said. "You’re checking behavior, not just execution."
Healthy Skepticism Meets Healthy Patterns
Matt summed it up: "You're not selling snake oil. You're selling tools. And tools I can debug, extend, and trust? That’s a win."
Shoutout to Matt for being a good sport and asking us the tough questions.