Agents need boring infrastructure

The impressive part of an agent demo is usually the least important part. A model writes a decent response, calls a tool, maybe edits something. That is the part everyone can see. The harder question is whether the product still feels trustworthy after a refresh, a failed tool call, a slow stream, or a half-finished task.

The agent work I like most is the infrastructure around the intelligence: streaming that does not drop turns, persistence that makes the conversation inspectable, replay buffers that let the UI recover, scoped tools that cannot wander outside the job, and evals that turn vibes into something you can regress against.

A useful agent is not just a chatbot with access to APIs. It is a product surface with memory, permissions, failure states, and a clear model of what it is allowed to do. The boring pieces are what let users build confidence in it.

That is also where product taste shows up. When should the agent ask a clarifying question? When should it take action? When should it show its work? When should it stop? The answer is different for a dispatcher, a technician, a customer success person, or a recruiter.

I am increasingly convinced that the agent itself is only one layer. The real product is the surrounding system that makes the agent legible, recoverable, and useful in the middle of real work.