|
Yesterday, there was a turning moment in the AI industry, as Anthropic cut its subscriptions to power third-party agentic harnesses.
This is a critical moment for several reasons, but it might also determine a weird inversion.
No Free AI Lunch | ||||||
| ||||||
 | ||||||
As we speak, today, April 4, 2026, Anthropic ended something that was never supposed to exist in the first place. | ||||||
|
We’re now seeing the “shape of the harness” form.
Indeed, the debate about which AI model is smarter is the wrong debate.
It has always been the wrong debate. Model capability matters at the margin — for the specific class of tasks where quality differential is large enough to justify premium pricing. But the structural question that determines who captures value in AI is not which model reasons better. It is who controls the layer through which model capability reaches the world.
That layer is the harness. And the battle over whether it should be open or closed is the most consequential architectural decision in AI right now — one whose outcome will determine competitive positions for years, not months.
 |
What a Harness Actually Is — and Why It Matters
 |
A harness is the orchestration infrastructure that links a user’s intent to a model’s execution. It decides which model gets called for which task, how context is structured and passed, how output is parsed and acted upon, how sessions are managed across turns, and how cost is controlled across workflows.
In a simple interaction — a human typing a prompt into a chat interface — the harness is invisible. The model and the interface are effectively the same product. But in agentic workflows — multi-step tasks, tool-using agents, automated pipelines, programmatic loops — the harness is where almost all the interesting architectural decisions live.
The harness is the intelligence layer above the model. It determines what the model sees, in what order, with what context, and at what cost. A well-designed harness running a mediocre model often outperforms a poorly designed harness running a superior model. This is not a theoretical claim — any developer who has spent serious time on production agentic systems knows it from experience.
This is why the harness layer is where the real competitive battle is being fought. Not at the model level, where quality is expensive to build and difficult to sustain. At the orchestration level, where architectural decisions compound into durable advantages.
The Two Architectures — Drawn Cleanly
 |
There are two coherent bets on how the harness layer should be structured, and they are structurally incompatible.
The closed vertical stack integrates the harness with the model la