Most teams with agents in production still improve them by hand: a failure comes in, someone edits a prompt or a tool description, and the fix quietly breaks something else. Machine learning was built for exactly this kind of problem, and an agent’s harness (the prompts, tools, and workflow around the model) has everything it needs to learn from data. So why are so few teams training it?

In this talk, we’ll walk through what makes harness optimization harder than it looks, what we learned from comparing Claude Code in a loop and other popular approaches on a real production agent, and what it takes to make it work. We’ll close with where this is heading: model-harness co-training, which optimizes the model and harness together.

Main Takeaways

1. Why agent harnesses should be trained and how to tell whether yours needs it.
2. How well Claude Code in a loop optimizes a production agent, and where it stalls.
3. Where agent optimization is heading, why optimizing the model and harness one after the other falls short, and how model-harness co-training fits in.