← All explainers
Explainer

Harness: The Software That Turns a Model's Answers Into Work

3 min read Reviewed
A small machine labelled 'Model' reads a page inside an open wooden booth labelled 'Harness', beside a tray of pages and a bell tagged 'Needs approval'. An arm fixed to the booth, not to the machine, pulls a folder from a filing cabinet beside it.

Key takeaways

  • A harness is the software that manages how an AI model works inside a system: it prepares what the model sees, handles requests to use the available tools, and applies any limits set for it.
  • The model determines what an AI system can understand and reason about. The harness determines what information it receives, what actions it can take and how those actions are controlled.
  • A limit holds only where software enforces it. An instruction to the model isn't such a limit.
  • When you compare AI tools, look at the harness as well as the model: what the tool can see, what it can reach, what needs approval, and what it records.

The software that runs around a model

Picture two AI assistants that run on exactly the same model. One answers questions about late orders. The other looks the order up, alerts the warehouse and asks a manager to approve a refund. The difference is the harness: the software that manages how an AI model works inside a system. It prepares what the model sees, handles requests to use the available tools, and applies any limits set for it.

On its own, the model reads what it’s given and writes a reply. That reply can be an answer, or a request to use a tool, meaning a function the system can run on its behalf, such as “look up order 4471”. The model can act only through the tools the harness makes available. The software that turns its replies into work getting done is the harness. Put simply, it’s the software around the model.

From a chat window to a working loop

Used broadly, as in this explainer, the word covers even a plain chatbot: the instructions the app gives the model, the conversation and the chat window. Some people use it only for AI agents, systems that work through a task in steps and use tools.

In an agent, the harness runs a loop. It calls the model with what it has prepared. The model replies with a final answer or a tool request. An allowed request runs, and the relevant result goes into the next call. This repeats until the model gives a final answer or the run is stopped.

A harness manages how the model works. A simple agent loop, as a simplified example. A large frame labelled Harness holds the steps. At the top of the frame, two dashed boxes for memory and a trigger are marked as optional. Below them, the harness prepares what the model sees: instructions, the conversation, a list of the tools it can use and the results so far. An arrow leads to the model, which replies in one of two ways: a final answer, or a tool request. The tool request reaches one check, highlighted as an example. A request that needs approval waits for a person, who approves or refuses it. The harness runs an allowed request's tool on a computer, a cloud or an isolated space, still inside the frame. The tool sends a request to an order system outside the frame, a separate system in this example, and gets a response. A line brings the relevant result back up, to be prepared for the next call. Notes say that the frame is the harness, the software around the model, and where it runs the tools. Blue marks the model, which the harness calls, and the highlighted check applies only to the requests that pass through it. The loop repeats until the model gives a final answer or the run is stopped. A harness manages how the model works. A simple agent loop. Simplified example. Harness Memory Trigger Optional, in any combination. Prepares what the model sees Instructions, chat, tool list, results The model replies Final answer Tool request A check, in this example Needs approval Waits for a person Approve or refuse Allowed The harness runs the tool Computer, cloud or isolated space Result back Request Response Order system Separate system in this example The frame is the harness: the software around the model, and where it runs the tools. Blue: the model, which the harness calls. Orange: one check, as an example. It applies only to the requests that pass through it. The loop repeats until the model gives a final answer, or the run is stopped.
One pass of a simple agent loop. The model replies with a final answer or a tool request. The harness prepares what the model sees, checks the request, and runs the tool, which here connects to a separate order system.

What a harness holds depends on the task:

  • Information: what the model sees, chosen to fit the context window, the most a model can handle in one call.
  • Tools: what the model can ask to use, and the environment they run in.
  • Memory: what’s kept between runs.
  • Triggers: a schedule or an email that starts a run.
  • Approvals: which actions wait for a person.

Think of the model as a skilled new employee who can work only through the company’s systems. They see only what’s put in front of them, act only through the tools they’ve been given, need a sign-off for big decisions, and know last week only if it was written down. The harness is that workplace.

Not the model, and not a guarantee

The harness isn’t the model. Changing it doesn’t retrain the model, and it doesn’t guarantee a correct result: the choice of model still shapes how well the system reasons.

Writing “never issue refunds” into the model’s instructions can’t guarantee that a refund tool won’t pay out. A limit holds only where software enforces it, so sensitive limits should also be enforced in the tools and systems themselves.

Late orders at a distributor

Back to the two assistants, as a hypothetical. A mid-sized distributor runs both on the same model. The first works only from what staff paste into the chat. The second can reach the order system and the warehouse, and any refund above a set amount waits for a manager, with the refund system itself refusing to pay before that approval. What differs is the harness: what each assistant sees, what it can reach, and what needs approval.

Judging a tool by more than its model

The model determines what an AI system can understand and reason about. The harness determines what information it receives, what actions it can take and how those actions are controlled. That doesn’t make the harness more important than the model. It means the model’s name alone can’t tell you what a tool will actually do in your business.

Our recommendation: before you trust an AI tool with real work, ask about its harness as carefully as its model. The harness is also where a team’s improvements can build up over time, one theme of where the harness pays off.

Questions for whoever owns the tool

  1. Besides the model, what can this tool see and reach?
  2. Which actions need a person’s approval, and what enforces that beyond the instructions?
  3. What does it keep between runs, and what can start it?
  4. If it goes wrong, can we see which steps it took and which tools ran?

Questions this explainer gets

Is the harness the same thing as the AI model?

No. The model reads what it's given and writes a reply. The harness is the software that prepares what it sees, handles its requests to use tools and applies any limits. Changing the harness doesn't retrain the model. It changes what the model sees, what it can reach and what needs approval.

Does a chatbot have a harness, or only an AI agent?

Used broadly, as in this explainer, a chatbot has a small one: its instructions, the conversation it sends along and the chat window. Some people use the word only for the software that runs AI agents, which adds tools and a loop, and can include memory, triggers and approval steps.