Already 7 months old when the scout caught it.
Harness engineering
What it is
Harness engineering is the discipline of designing everything around a model that turns raw model intelligence into a reliable, autonomous agent: the environment it runs in, the tools and permissions it has access to, the feedback loops and observability that catch its mistakes, the docs and rules it's steered by, and the guardrails that keep it from repeating a failure. It's usually shorthanded as the formula "Agent = Model + Harness" — the model supplies the intelligence, the harness is everything else that makes that intelligence usable and safe to run unattended.
Why it's emerging now
As models plateaued on raw capability gains, practitioners found that the bigger lever for making agents actually reliable in production was the scaffolding around the model rather than the model itself. OpenAI made the term concrete on February 11, 2026, describing how a team shipped a large internal product with zero hand-written code by engineering the agent's environment rather than its prompts ("humans steer, agents execute"). Anthropic followed with its own engineering posts on multi-agent harnesses for frontend design and long-running autonomous software work. The idea kept spreading through the year and was called out as one of the defining trends of AI Engineer World's Fair 2026, alongside a shorthand that stuck: "if you're not the model, you're the harness."
How it relates to what you know
This is the formalized discipline built on top of "agent harnesses" already on your ledger — the same relationship "context engineering" has to context windows, or "loop engineering" and "graph engineering" have to agent execution topology. Where agent harnesses is the noun (the wrapper around the model), harness engineering is the named practice of deliberately building and iterating on that wrapper, with its own guides, playbooks, and now a body of case studies from labs and practitioners.