The First and Last Harness Is Us
What agent harnesses are, why they matter, and why they won't disappear.
Batman has no superpowers. He's just the smartest person in every room he walks into, and that's not enough. What makes him dangerous is the belt: the grapnel, the gadgets, the Batcomputer back home, and Alfred in his ear. Take the belt away and you have a brilliant man standing on a rooftop with nothing to do.
Today's AI models have the opposite problem from Batman, and the same solution. They're close to superhuman at thinking. They have read more than any person ever will, and they can reason through problems that would take you an afternoon. But on their own they have no hands, no memory, and no idea where they are. Picture hiring a genius and sitting them at an empty desk: no laptop, no logins, no notes, nobody telling them what the job is. They'd be brilliant, and useless.
The stuff on the desk has a name. It's called a harness. Every useful AI agent you've heard of is a model plus a harness, and the harness is quietly becoming the most important part of the story.
From autocomplete to agent
It helps to see how we got here, because each step added something the model didn't come with.
Autocomplete. At the bottom, a language model does one thing: it predicts the next word. Feed it a sentence and it guesses what comes next, over and over. Trained on enough text, that simple trick turns into something that looks a lot like knowledge. But a raw model is like a library that only speaks when you start a sentence for it.
Assistant. Next, labs trained models on human feedback so they would follow instructions and hold a conversation. That's the ChatGPT moment. The genius could finally talk to you.
Tools. Then models learned to ask for help. Instead of guessing today's weather, the model writes a request to a weather service and reads the answer back. Search, calculators, databases, your calendar. The genius got a phone.
Reasoning. Newer models think before they answer, working through a problem step by step and spending more effort on harder questions. The genius got scratch paper.
Agents. Put all of that in a loop and something new happens. The model makes a plan, takes an action, looks at what happened, and decides what to do next, again and again, until the job is done. That loop is what we mean by an agent. It doesn't just answer. It works.
So what is a harness?
A harness is everything wrapped around the model that lets it do real work. Think back to the genius at the desk. A harness gives them:
An onboarding binder: the instructions that tell the model who it is, what the job is, and how to behave. A laptop: the tools it can use, from search to code to clicking around a website. A notebook: memory, so it remembers what you told it last week and can look things up in your files. An inbox tray: the loop that keeps work coming in and results going out. A keycard: permissions, meaning what it's allowed to touch, spend, or send, and when it has to ask you first.
Why do we need any of this? Because a model on its own forgets everything the moment a conversation ends. It can't see your files. It can't press a button. And when it's wrong, it's confidently wrong, with nothing around it to check. The harness is what turns raw intelligence into something you can trust with a task.