We’ve learned to continuously improve our software. We fix bugs, run retros, review incidents, and turn recurring problems into changes. But when it comes to the systems we build around our agents, we often treat them as finished once they work.

Until the same annoyance happens for the fifth time.

You ask for an integration test and get a unit test. You correct the agent, adjust an instruction, and move on. A week later, you have the same conversation again. The harness works, but it isn’t getting better.

This talk explores what happens when we apply the principles of continuous improvement to the agent harness itself. Instead of treating repeated friction as something to work around, we’ll look at how those interactions can become signals that tell us where the harness needs to change.

Along the way, we’ll explore practical ways to discover these signals, turn them into deliberate improvements, and create a feedback loop between how we use our agents and how we build the systems around them.

Because the goal isn't to write the perfect harness. It’s to build one that gets better as you use it.