Agent Harness
The wrapper of code around the model that turns raw text prediction into an agent that can actually do things.
Think of it like
A climbing harness — the rope, clips, and straps that let a capable climber attempt something they couldn’t do bare-handed.
Example
The same model gets 20% on a coding benchmark with a bare prompt and 60% once you put it in a good harness that lets it run tests and read errors.
How it actually works
A model on its own only predicts text; the harness is what gives it tools, feeds back results, and lets it retry. A surprising amount of “model progress” on agent benchmarks is really harness progress — better scaffolding around the same weights. That makes it hard to compare models fairly, since a strong harness can flatter a weak model.
For product teams
When comparing agent results, ask how much is the model and how much is the scaffolding around it.
For engineers
The harness supplies tool interfaces, feedback loops, and retry logic that wrap the base model into an agent.
Related
- Agent Loop — The harness runs the think-act-observe cycle.
- Agent Evaluation — How much of the score is model vs. scaffolding.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome