What is an AI harness?: Getting LLMs to do work well
Published: October 8, 2026 · 1 min read
A model on its own can only do so much. What it gets done depends a lot on what's around it: the context it has, the tools it can use, the instructions it's given, and whether it can check its own work.
That surrounding layer is called a harness. A few people have defined it well.
LangChain puts it simply: "A harness is every piece of code, configuration, and execution logic that isn't the model itself." (The Anatomy of an Agent Harness)
Birgitta Böckeler calls it "shorthand to mean everything in an AI agent except the model itself." A good one, she says, makes the agent more likely to get it right the first time and gives it a feedback loop to fix its own mistakes. (Harness engineering for coding agent users)
It was also the whole topic of a recent YC Paper Club episode, Why The Harness Matters More Than The Model. The description makes a pretty bold claim: "The same model weights that score 30% on ARC-AGI score 95% with a better harness."
Diagram
Model, harness, and agent
Agent = Model + Harness
AGENTS.md is a simple example. It's a file in a repo that tells an agent how the project works, like where things are and how to run the tests. It's basically an onboarding doc for the agent. But a harness can go well beyond a markdown file. It can shape what an agent can access, what it can change, how it gets feedback, and how it knows when it's done.
Diagram
What a harness adds around a model
Harness
Set up before it starts
Instructions
How to work, what good looks like
Context
Files, history, memory
Permissions
What it can access and change
Model
Reasons about the task and decides what to do next
The loop: act, see what happened, adjust
Tools
Search, edit, run commands
Feedback
Test results, errors, review
Swap any of these parts and the same model behaves like a different system.
Without one, an agent is mostly working blind. It has no real context to go on, no way to keep track of where it is in a long task, no standard way of doing the work, no way to check whether it worked, and no limits on what it can touch.
If you want to go deeper, awesome-agent-harness is a curated list of hundreds of harness projects and readings, and the survey Agent Harness Engineering: A Survey (Li et al., 2026) goes into the research side.
You don't necessarily need a better model to make the system better. You can take the same model, change the tools, context, instructions, or feedback loop, and it can do something it couldn't before.
We've spent a lot of time asking which model is best. I think we'll spend a lot more time asking what we've built around it.
The model matters, but the environment you give it might matter just as much.