Explaining AI harnesses: model vs. tools vs. agent
A primer breaking down LLM, harness, agent, interface and MCP, explaining why assistants like Grok Bot feel different on the same model.
if you’ve heard people talk about “AI harnesses” and still aren’t totally sure what that means, read this. the simplest way I think about it: LLM = generates the text harness = gives the model tools, tells it how to call them, runs those calls + sends the results back agent = LLM + harness interface = the thing you interact with MCP = a way for the harness to reach tools + data living somewhere else so Claude Code, Codex, Cursor, Grok Bot, OpenClaw, etc. can feel VERY different even when the underlying model is the same. the model is only one very small piece of what you’re really using.
This post quotes another post: x.com/theo/status/2113424203345265053