A harness is the thing that turns sloppotron spew into action. That's... pretty much it. LLMs are still "just" token prediction algorithms, but you can coax the

A harness is the thing that turns sloppotron spew into action. That's... pretty much it. LLMs are still "just" token prediction algorithms, but you can coax them into outputting things that look like commands. The harness figures out which bits are supposed to be commands, executes them, and feeds the output back into the context of the LLM. They also have prompts of their own to guide LLM behavior.