What Is an Agent Harness?

A chatbot tells you how to do something. An agent goes and does it. That is the whole difference.

People call three different things AI. A model is text in, text out, with no memory and no hands. A chatbot is a model plus a conversation, but it is still just talking. An agent takes an action, sees what happened, and decides the next action. Then it repeats.

Every agent, underneath the branding, runs one loop. The model decides the next action. The harness runs that action. The result goes back to the model. The model decides again. Think, act, observe, think again, until the goal is met. On a long job the model also keeps a plan it can revise as it learns.

The loop needs something to run it. That something is the agentic harness, and it is the part people leave out.

Here is a clean way to draw the line. The harness is every piece of code, config, and infrastructure that is not the model itself. A raw model just takes in text and puts out text. The harness is what lets it run commands, edit files, and remember things.

It runs the loop. It gives the model a filesystem, a workspace to read and write and offload work that outlives one session, with git to track it. It gives the model a computer, a bash tool in a sandbox, so it can write its own tools instead of waiting for pre-built ones. It fights context rot, by compacting a full window, sending huge outputs to a file, and loading skills only when needed. It adds knowledge the model was never trained on, through memory files and web search. It delegates, spawning subagents with their own context for parts of the job. And it holds the guardrails: step limits, approvals, blocked commands, and recovery when a call fails.

Watch one run. The goal: fix the broken build. The model decides to look at the last commit. The harness shows it. The model spots a renamed function with an old caller. The harness runs a search. The model edits both files. The harness applies them. The model runs the build. It passes. Nine decisions by the model, nine actions by the harness.

Where does thinking mode fit? The "decide the next action" step is often a thinking step. The model reasons privately about what it has learned, then commits. So an agent is really a thinking model, in a loop, with hands.

The hype says the agent is the intelligence. But the harness is where the leverage is. On the Terminal-Bench benchmark, the same model with a stock harness ranked around thirtieth. Swap in a harness tuned for the task, same model weights, and it jumps to the top five. Intelligence lives in the model. Usefulness comes from the harness around it. Pull the harness out and you have a model describing what it would do, to nobody.