What is harness engineering?

Harness engineering is the discipline of designing and evolving the harness around a model, the loop, tools, context management, safety controls, orchestration, and extension surfaces, so an agent’s mistakes get fixed at the system level instead of waited out. A source-code study of eleven production coding harnesses, which named the discipline in early 2026, maps it to seven canonical subsystems and catalogs 29 recurring design patterns across them.

In practice it’s narrower than architecture: Addy Osmani describes the job as noticing when an agent makes a mistake and engineering a fix so it never makes that exact mistake again, a prompt line added, a tool description tightened, a hook inserted, rather than waiting on a better model. The scope stretches past the loop itself into how an agent’s actions get observed, since a harness engineer can’t fix what a trace doesn’t show.

It never finishes: a fix that’s correct today still needs the next one once the model, the tools, or the traffic underneath it changes.