Why this blog exists
Demos of AI agents are everywhere. Notes from running them in production — with real users, real third-party APIs, and real token bills — are rare.
I build agent tool integrations for a living. Most of what I've learned came from failures that no documentation warned me about: tools that models refuse to pick, schemas that quietly burn thousands of tokens per request, connectors that pass every compliance check and still confuse every model that touches them.
This blog is where I write those down. Expect posts about MCP servers, tool design for LLMs, eval harnesses, and the economics of agent systems.
First real post coming soon. It involves data, and the data is not flattering.