Harness

Agent harness

In one sentenceThe software wrapped around an LLM that controls it: what it sees, which tools it can use, what it remembers and when it stops. The model is the horse; the harness steers it.

What it means

Think of a horse. The horse is powerful, but on its own it just wanders the field. The harness is what connects that power to the cart, and the reins are how you steer it. An LLM is the same: raw intelligence with no hands, no memory and no direction. The harness gives it all three.

A harness is everything around the model: the System Prompt and instructions, the tools it is allowed to call (Tool Use, MCP connectors, a browser, a terminal), its Memory, the loop that lets it take a step, check the result and take the next step, the permissions that say what needs your approval, and the channels you talk to it through (an app, WhatsApp, Telegram, Slack).

Claude Code, OpenAI's Codex, the ChatGPT app, OpenClaw, Hermes Agent and Grok Bot are all harnesses. Many of them can run different models inside. That is why the same model can feel brilliant in one product and clumsy in another: the harness makes a huge difference.

How to use it

• When choosing an AI agent, judge the harness, not just the model: what tools can it use, what does it remember, where can you talk to it, and what can it do without asking you?
• Set the reins before you let it run: decide what it may do alone (research, drafts) and what always needs your OK (sending, paying, deleting).
• If an agent keeps making the same mistake, fix the harness (its instructions, skills or tools), not just the one answer.
• Swap the horse, keep the harness: many harnesses let you change the model underneath, often through OpenRouter, to trade cost against quality.

I want an AI agent to [job, e.g. research new leads every morning and draft intros]. Design the harness for me: the instructions it needs, the tools and accounts it should connect to, what it should remember, what it may do on its own, and what must wait for my approval.

Related terms

AI Agent OpenClaw Hermes Agent Grok Bot OpenRouter LLM Human in the Loop