LLM vs Agent vs Harness, Explained by a Caveman

Everyone argues about AI agents. Half the people arguing can't say where the LLM ends and the agent begins.

So here's the whole thing, explained by a caveman.

The LLM: a big brain in a jar

The LLM is a brain. A very big, very well-read brain. It knows many words. It can finish your sentences, write your emails, and explain recursion five different ways.

It has no hands.

You give it a question, it gives you an answer. That's the entire product: text in, text out. It doesn't remember yesterday unless you hand it the notes. It can't open a browser, run your tests, or move your money.

The agent: a brain with hands

Give that brain hands, a goal, and permission to keep going.

An agent takes "book me the cheapest flight to Dubai next month," breaks it into steps, searches flights, compares prices, hits a dead end, tries different dates, and comes back with a booking. Nobody held its hand between steps. The loop did that: think, act, look at what happened, think again.

The LLM is the part that reasons. The agent is the loop around it: tools it can call, memory of what it already tried, and the stubbornness to keep going until the job is done.

The harness: the leash

The harness is everything around the agent that keeps it from running off a cliff. Which tools it's allowed to touch. How many times it may retry. What it must never do. Where it writes logs. How you score its work.

Skip the harness and your agent is a toddler with a credit card. It will do something. It won't be what you wanted.

Building agents for production taught me this split: the model is maybe 10% of the work. The harness is the other 90%. Anyone can call an API. The work is guardrails, evals, retry budgets, and permissions.

The mammoth stew test

The village wants mammoth stew.

The LLM hands you five recipes and wishes you luck.

The agent goes out, hunts the mammoth, and brings back meat.

The harness says no baby mammoths, home before dark, max three spears.

Same goal. Three very different cavemen.

When to use which

Need text, a summary, or an answer? A single LLM call is enough. Don't build an agent because it's fashionable.

Need a multi-step job done with tools? That's an agent.

Letting it touch anything real? Build the harness first. Permissions before autonomy.

The short version

LLM = brain. Agent = brain + hands. Harness = leash.

Next time someone says "we built an AI agent," ask what the harness looks like. The answer tells you whether it's a demo or a product.

Which one are you building right now?