Half Baked Newsletter April 8th.jpg

The Idea: In early 2025, Aryan Shah, Aayush Sheth, and Marcus Yearwood spotted a problem nobody was talking about. Everyone was building AI agents. Nobody was making them actually work. The gap wasn't the models - it was the layer underneath them. Enterprise agents kept breaking in production because they had no way to learn from real workflows, no evaluation framework to catch failures, and no system to improve over time. They were brittle by design. Aryan, Aayush, and Marcus decided to build the missing layer - a post-training and continual-learning infrastructure that turns a company's own data, tools, and environments into an agent stack that actually gets better the longer it runs. Their first product, Insight, generated training signals and ran reinforcement learning loops so agents could pick the right tool, preserve context, and complete long multi-step workflows reliably at scale. They called the company Metis - after the Greek goddess of wisdom and skill. This was the beginning of Metis.

The Execution:

The lesson? The best infrastructure companies don't fight for attention - they get quietly embedded into a customer's core workflows until the customer can't imagine operating without them. Aryan, Aayush, and Marcus didn't try to build the flashiest agent demo. They built the layer that made everyone else's agents actually work. When your customer starts thinking "we should just own this" - that's not a red flag. That's the exit.