The latest

Companies are deciding whether to build or buy the software that controls how artificial-intelligence models use corporate data, execute tasks and consume tokens. Known as an AI model harness, this layer wraps around a model to provide memory and business context while enabling AI agents to connect with systems, run code and manage workflows. The choice can affect companies’ privacy, resilience, infrastructure investment and reliance on individual AI providers.

Details

• Control layer: David Pan, a director and AI industry practice lead at Moody’s, describes the harness as a system surrounding the “brains” represented by AI models. Much as a physical harness lets a rider guide a horse, the software allows users to direct models. Combined with a capable reasoning model, it can link the model to operational systems and organize the steps needed to complete work.
• In-house resilience: Pan calls the development of proprietary software around models “harness engineering.” Separating workflows from the underlying models allows businesses to switch models and become less reliant on one provider. Bringing the harness in-house can therefore strengthen business resilience. Pan argues that organizations in regulated sectors, including banking and government, should build their own systems to keep workflows private.
• Vendor alternative: OpenAI and Anthropic offer model harnesses to enterprise customers, giving businesses an alternative to developing the surrounding infrastructure themselves. Bristol-Myers Squibb selected Anthropic’s Claude as its “standard harness.” Greg Meyers, the biopharmaceutical company’s chief digital and technology officer, said that approach avoids rebuilding basic infrastructure tooling already available from the vendor.
• Cost routing: A router is another critical harness component. The software can automatically select between frontier models and less expensive alternatives according to the task. Companies increasingly use this capability to limit token spending rather than deploying the most capable and costly model for every request, making model selection part of day-to-day cost control.
• Moody’s system: Moody’s built a harness-like tool called Research Assistant. The AI agent chatbot can use different Moody’s datasets and switch among AI models on the back end, Pan said. The architecture gives the credit-ratings and research company a way to retain its own data environment and workflow while drawing on more than one model.

Background

The model-harness concept became more prominent in 2026 alongside the development of reasoning models. Terminology continues to vary, encompassing prompt engineering, context engineering and harness engineering. Pan said the durable requirement across those approaches is supplying language models with the right context, particularly the business data needed to perform useful work.

What’s next

The next indicator will be companies’ procurement and development decisions: whether they bring harnesses in-house, adopt vendor systems or combine the two. Their deployment of routers across production workflows will also show how aggressively they are balancing model capability against token costs.