For most of computing history, applications encoded behavior while operating systems abstracted the hardware beneath them. That separation gave us portability, composability, and the modern software industry.

AI agents now generate behavior at runtime, so software needs a new layer to make that behavior useful, reliable, and repeatable.

That layer is the harness.

A harness gives a model a stable operating environment: tasks, context, tools, persistent state, permissions, feedback loops, and a definition of done. It determines what the agent can see and change, how it interacts with the world, and how its work is tested and recovered. The model supplies capability. The harness turns that capability into dependable software.

Using coding agents as a case study, this keynote builds a first-principles model for harnesses and maps familiar operating-system concepts like lifecycle, resource management, memory, I/O, protection, and observability onto agentic systems.

You will leave understanding how harnesses work, how they compose, and how to design better agents by building better harnesses around them.