You Have a Harness: What to Keep, Test, and Improve
Understand how to design and refine a harness that controls AI agent behavior by establishing boundaries, verifying outcomes, and preserving session evidence. Learn to build the smallest effective harness loop, manage scope contracts, verify claims, and iteratively improve reliability based on observed failures to maintain dependable agent work.
You have finished the course. You started with an agent that could produce a plausible patch and built toward a system that can establish what work belongs in a run, limit what the agent may do, prove what changed, and leave an honest record for the next session.
That is harness engineering. It is not simply a longer or more detailed prompt. It is the practical work of making agent execution inspectable, bounded, and recoverable.
We built not one control, but a system of controls that answer different questions at different times.
The course’s core idea is that capability and reliability are different jobs. A capable model notices a problem, ...