Adam DeJans Jr. wrote: “A harness is the software environment that gives an LLM the context, tools, state, evaluation criteria, and feedback loops required to actually perform work. Instead of a simple prompt → answer interaction, the system can understand an objective → inspect its environment → take an action → observe the result → evaluate it → modify its approach → repeat.

That difference is enormous. The exact same LLM can produce radically different levels of useful work depending on what has been built around it.” Link