The harness is the product

In February 2026 the phrase 'harness engineering' left the labs. OpenAI's Ryan Lopopolo used it. Mitchell Hashimoto argued for it. Practitioners had already been living it: on SWE-bench Pro, swapping the agent harness moved pass@1 more than many model upgrades. That sentence is the founding fact of this market.

A harness is not a prompt. It is the loop that offers tools, the permission system that decides what those tools may touch, the memory that survives a session, the recovery that turns a failed call into a retry instead of a bill, and the stop condition that a model will not apply to itself. Think of the model as a CPU, the context window as RAM, and the harness as the operating system. Nobody prices an OS like a CPU cycle. The 2024–2025 market did, which is why so many 'agent companies' were model wrappers with a landing page.

The commercial pattern that remains is unromantic. Open source with a paid host. A vertical harness for a regulated industry. A marketplace of skills that the harness knows how to load. A control plane that will still answer the phone at ten thousand agents. Harnessmarket is the last two, and the registry that makes the first two tradable.

Three implications follow. First, procurement will move from 'which model' to 'which harness, which mix, which license'. Second, evaluation has to hold the model constant — Lattice's SWE Eval Pack exists because almost nobody does. Third, attribution is no longer a nice dashboard. If two harnesses touched the same ticket, legal, finance, and engineering all need a number. Token share is not that number.