Ryan Lopopolo

An agent is a parameterized program over a set of capabilities: a model, context, a computer, a filesystem, tools, runtimes, network access, identity, guardrails, and the instructions that bind them together.

We are beginning to recognize those primitives. What we do not know is the best implementation of almost any of them.

That makes the architecture of an agent platform unusually important. If the platform hard-codes today’s assumptions, every new idea becomes a feature request. Instead, it should behave more like a microkernel: expose composable primitives, make their implementations swappable, and let builders invent the agent they need.

This talk develops that model from first principles and shows how harness engineering, tool discovery, verification, governance, judges, and autonomous software factories emerge naturally from it.

We are early enough that we should optimize for discovering what works, not standardizing what happened to work first.

Reasoning models and coding agents have changed what is possible in software development, but most teams are still using workflows built for a world where humans write nearly all the code directly. That mismatch is where much of the disappointment around AI coding comes from.

This talk introduces Harness Engineering: an agent-first approach to building software where the key engineering task shifts from manually producing every implementation detail to designing the harnesses that let agents do useful work safely, repeatedly, and at high quality.

I’ll share practical patterns for making AI coding actually work in production: how to scope autonomy, shape context, design eval and verification loops, structure tool access, and build feedback systems that keep agents effective without making them useless. I’ll also cover where agents are already strong, where they still fail, and why “just give the model the repo” is not a serious deployment strategy.

If you are trying to move from demos to real delivery, this talk offers a concrete framework for how agent-first engineering actually works.