Harness Engineering: Core Philosophy

Everything in this series hangs on one move: stop treating the model as the agent, and start treating the agent as a system. Zhong and Zhu write it as a rough equation, where a system’s capability is a function of the model, the harness, the environment, and the task:

C_system = F(C_model, C_harness, C_environment, T)

Read it plainly and it says the model is one term of four. A capable model with a bad harness is a brain with no hands. A modest model with a great harness can be dependable enough to trust with real work.

From claim to evidence

The other half of the philosophy is about proof. In a good harness, “done” is not something the model asserts. It is an evidentiary object, produced by the system and checkable by a human. That single shift, from claim to evidence, is what separates a demo from something you would let touch production.

The autonomy gap

There is a name for the thing a harness closes. The autonomy gap is the distance between what a model can do in the small, a function here, a test there, and what it can actually finish on its own without a human stepping in.

Every time a human has to intervene, it points at a missing part of the harness. Zhong and Zhu call that signal missing-harness human intervention, and it is diagnostic: it tells you exactly which runtime support was absent when the agent got stuck.