Agent harness

  • Agents
  • Standard term
7 steps

An agent harness is the software around an AI model that runs its work. It supplies information and tools, records steps, and handles replies and errors. It can also enforce permissions and other limits.

A model proposes responses or tool requests. The harness handles each request under its rules, runs an allowed tool, and passes the result back to the model. That loop continues until completion or a stop condition, so tool access and recovery behavior depend on the surrounding software.

Where you’ll meet it

An AI company posts that its coding agent now finishes far more tasks on a public coding test than it did last month, using the same model underneath. The post credits a new harness that runs the project's tests after every edit and tries a different approach when one fails.

What to do

Read the gain as news about that one product. Another app built on the same model may not match it. When you pick a tool for your own work, give two candidates the same task you know well, and compare what each one does after its first mistake.

Easy to mix up

An AI agent is the whole system that works toward a goal, model and harness together. The harness is the software around the model that supplies tools, tracks steps, and enforces limits.

Check yourself

While tidying up, your coding agent deleted a file outside the project folder. Which change directly enforces that folder boundary?

Go deeperHarness controls bring tradeoffs as well as benefits

Researchers often call the software around a model a “scaffold.” In agent benchmarks such as SWE-bench, which uses real issues from open-source projects, a score reflects a model working within a particular scaffold. Holding the scaffold fixed can help isolate a model change; changing both compares two complete systems.

Retrying can recover from a dropped connection. If a purchase succeeded and only its confirmation was lost, however, another attempt may duplicate the charge unless the service prevents repeated effects. Tight permissions can contain damage while also blocking a legitimate task that needs access beyond the allowed scope.

On a long task, the history can fill the context window, the text the model can take in at once. Some harnesses summarize older steps to make room. Those summaries may omit details, so important state needs a way to remain available beyond the compressed conversation.

Practice Modern AI

Or take the learning quiz on a mix of every set