Infrastructure layer
The Harness
The execution runtime around the model, hardened for regulated deployment. It decides which model runs, what it is allowed to touch, what happens when it goes wrong, and what you can see afterwards.
An agent is a model plus a harness
The harness is everything wrapped around the model: the prompts and rules that define what it must never do, the tools and integrations that define what it can actually do, the files and memory that let it keep its work, the loops and planning that carry a task to completion, the checks that catch mistakes live, and the sandbox it runs inside.
The evidence that this is where performance lives is now public. LangChain moved their coding agent from 52.8 to 66.5 on Terminal Bench 2.0, a 13.7 point improvement, by changing only the harness while holding the model fixed.
Source: LangChain, Improving Deep Agents with harness engineering, 2026.
Model routing without lock-in
The Harness routes across models rather than depending on any single lab. Each use case carries hard constraints on cost, quality and latency, and routing happens inside those bounds. A complaint response and an AML investigation have different tolerances, and they should not be served by the same model at the same price.
This also protects you commercially. When a better or cheaper model ships, it becomes a routing decision rather than a rebuild.
Model choice is becoming a live operating decision rather than a procurement one. Stanford’s 2026 snapshot put the top four labs within 22 rating points of one another, which means differentiation is shifting to cost, reliability, latency and domain fit.
Source: Stanford HAI AI Index 2026.
The economics
A well-engineered harness produces enterprise-grade output from smaller, cheaper models. Routing under explicit cost and quality constraints turns token spend into something a finance team can plan against, and the savings compound as volume grows.
The unit of measurement is moving with it. In EY’s July 2026 AI Pulse survey, 82 percent of senior leaders investing in AI said they were concerned about token usage and cost, and 98 percent of those using token-based AI said cost had caused them to reconsider some aspect of their approach. Only 64 percent said their organisation actively monitored token usage against clear budget guardrails.
Source: EY AI Pulse survey, July 2026.
We are measuring our own cost per governed decision across deployments before publishing figures. When we have numbers we can stand behind, they will appear here with their methodology.
Observability on your terms
Evaluation and observability are built around your KPIs and your rubrics, not a generic dashboard. If your complaints team measures fair outcomes and your clinical team measures information accuracy, those are the things the system reports against.