Harness engineering moves to the center of enterprise AI

Enterprise AI attention is shifting from frontier models toward the harnesses that surround them. A typical agentic harness includes instructions, memory, tool access, execution controls and a loop that keeps the model working until a task is complete, while enforcing permissions, workflows and governance.

Studies cited in the discussion say smaller models can reduce LLM operating costs by as much as 90 percent, and a Google playbook is described as putting 90 percent of an agentic system’s value in context or harness engineering, with the model accounting for 10 percent. The argument is that better harness design can reduce dependence on expensive frontier model calls, route work to smaller models and stop runaway agent loops before costs escalate.

Tokenomics is also becoming a sharper enterprise concern, with Palantir CEO Alex Karp criticizing OpenAI and Anthropic business models and emphasizing customer control over compute, models, data stacks and business advantage. Security risks remain unresolved, including examples where agent tools can be hijacked through trusted workflows.

For EU AI Act readiness, almost half (47%) of the diginomica network community has an AI tool and integration inventory with some gaps, a little over a quarter (27%) has partial visibility, and just under a quarter (23%) has fully mapped AI across the entire business.