The term "AI harness" is appearing more frequently as enterprises move from experimenting with individual AI agents to building workflows that combine multiple agents and models. Vendors are beginning to package the concept into products for enterprise customers, but the terminology remains unsettled. Questions about what qualifies as a harness, where it fits in an agentic workflow and when an enterprise actually needs one are still open.

In an interview with AI Business, Gareth de Bruyn, CEO, founder and chief architect at Debcor Engineering, which specializes in native AI integrations with SAP, discussed the emerging role of AI harnesses in enterprise workflows. He explained how harnesses coordinate multiple agents, manage interactions with enterprise systems and provide controls around AI-generated actions.

What is an AI harness?

De Bruyn said an AI harness sits between agentic code and the models themselves. It handles routing, access control, context management, evaluation and auditing. Those controls need to be established before an agent interacts with enterprise systems.

The value of agents becomes more apparent when multiple agents work together, he said. A harness makes it possible to coordinate those agents and manage their interactions.

Governing interactions with enterprise systems

A harness coordinates how an agent performs its work and interacts with enterprise systems. De Bruyn said that "transacts" may be a better description of that interaction.

An agent may access an enterprise system to gather information needed to make a decision. When it takes an action—such as creating an order, updating a customer record or moving goods—the harness applies the necessary guardrails and controls.

An agent may also need to interact with several enterprise systems to complete a task. The harness governs those interactions by helping determine what the agent can access, which actions it can take and how those actions are carried out.

An air traffic controller for agents

De Bruyn compared a harness to air traffic control. A sales order, for example, might arrive by email, fax or as a perfectly formatted order. The harness serves as the coordinating layer that evaluates the input and determines how it should proceed.

It might recognize the order, retrieve the relevant data and create the order. Alternatively, it might determine that the request must go through verification checks. Those checks could require several AI calls, increasing costs and creating additional auditing and tracking requirements.

The harness ensures that each request is routed appropriately and that the relevant checks are applied. The result is a standardized order that can then be entered into the customer system.

When does an enterprise need a harness?

For a single agent operating independently, De Bruyn said he would question whether a harness is necessary. He gave the example of a simple optical character recognition, or OCR, application that he created. The application allowed him to take a picture of a conference attendee's badge, have an agent read the information immediately and save it in his customer relationship management database.

That type of lightweight and straightforward task does not require a harness, he said. The value of a harness becomes clearer when multiple agents must work together or when their actions require greater oversight and coordination.

Build or buy?

Enterprises are also facing a build-versus-buy decision around AI. De Bruyn said many organizations are looking for pre-made AI solutions because they do not always have the expertise or confidence to build such systems themselves.

Pre-made harnesses and other packaged AI options are beginning to emerge, but De Bruyn said he does not consider the market mature. He also warned that the terminology can be misleading: a solution marketed as a single AI agent may actually involve many agents working together.

At SAP's Sapphire conference, for example, attendees heard about an accounts payable agent. De Bruyn said the workflow behind that offering could involve 10 to 15 different agents handling tasks such as invoice processing, information checks and routing decisions.

Although the product may be presented as one agent, multiple agents and systems can be working behind the scenes. In De Bruyn's view, marketing and operational reality have not fully caught up with one another.

More pre-made solutions are likely to emerge, he said, but enterprises need to understand what those products contain, how the different agents work together and what role the harness plays in coordinating them. For enterprises evaluating those offerings, the practical questions therefore extend beyond the product label to the systems it can access, the actions it can take and the checks it applies.

Start with the business outcome

De Bruyn's main lesson for enterprise leaders is that "AI is not magic." He said enterprises face significant pressure to adopt AI, similar to the pressure previously associated with technologies such as cloud computing and the Internet of Things, or IoT. Leaders should instead begin with basic business questions: What is the organization trying to achieve? What outcome does it want, and how will that outcome be measured?

After defining those goals and key performance indicators, or KPIs, an enterprise can determine what the AI system needs to do and where a harness fits into the design.

De Bruyn described a harness as a collection of components that brings together the models, agents and other systems required to deliver a specific outcome. He said organizations do not need to begin with the technical details. They should first define the business outcome, break down the required components and determine the role the harness should play.

At its core, he said, a harness helps manage risk and efficiency by coordinating those different components.

Editor's note: This interview has been edited for clarity and conciseness.