The Agent Harness: The Software Layer That Decides What AI Can Do

Two AI scheduling assistants can run on the identical model, yet one can only copy-edit meeting agendas while the other checks invitees' calendars, finds free time, and sends invitations. The difference sits entirely in the software that surrounds the model — the agent harness.

The Nielsen Norman Group defines the agent harness (also called an agentic harness, or simply a harness) as the software around an AI model that runs its execution loop. It carries out the model's requests to use tools, returns the results, and controls what information the model receives and what it is allowed to do.

What does the harness actually do?

A large language model generates data by predicting what comes next in a sequence. It can generate structured requests to check a calendar or send an invitation, but it cannot perform these actions on its own. To act, it must ask other applications to do things for it by outputting commands.

The harness handles each such request in three steps:

• it checks permissions
• it calls the appropriate tool
• it returns the result to the AI model

The model can then request another action, or return an output to the end user. This cycle of request, action, and result is the execution loop.

Why the loop matters more than the model

Running this loop is what transforms an AI model from a simple chatbot into an AI agent that can persistently work through a task, rather than just replying to a prompt. The harness, not the underlying model, determines whether a product answers questions or completes work.

For practitioners, the implications are direct:

• Capability is a harness decision. Context, permissions, and limits set in the harness define what users can accomplish, independently of model quality.
• The same model can power divergent products. Teams choosing between AI tools should evaluate the harness configuration — which tools are exposed, which data flows back — before comparing model benchmarks.
• Design work shifts to the loop. When an agent fails mid-task, the cause is often a permission boundary or a missing tool call in the harness, not the model's reasoning.

The comparison is straightforward: a model without a harness is a text predictor, while a model with a well-configured harness is a system that reads calendars, drafts invitations, and acts on a user's behalf.

What should teams watch next?

As the article argues, the harness's context, permissions, and limits shape what users can do. That framing suggests the next wave of product differentiation in AI interfaces will come less from model selection and more from how deliberately teams engineer the execution loop — which tools they connect, and where they draw the lines.