Harnyx — Vision Statement
AI industry is split on a central question: Big Model or Big Harness?
One camp — led by the foundation model labs — argues that scaffolding is temporary. That reasoning models will absorb most of what harnesses do today, reducing them to a minimal layer. As Noam Brown put it: “People are building scaffolding on top of the reasoning models right now. But I think in many ways, those scaffolds will also just be replaced.”
The other camp — the agent engineers, the systems builders — argues that harness engineering is where durable moats are built. Even as models become more capable and widely accessible, the teams that build better constraints, smarter orchestration, and tighter feedback loops around them compound their advantage over time. The model is the engine; the harness is the car. A more powerful engine helps everyone — but the team with the better-engineered car still wins the race.
We think both camps are missing something.
Deep research is a natural proving ground for this debate. It’s inherently a systems problem: query decomposition, parallel retrieval, source ranking, cross-reference verification, synthesis. No single model excels at every stage — the winning strategy requires a harness that orchestrates the entire workflow. And the optimal harness changes every time the underlying models change. A static harness gets outrun by the next model.
The real question isn’t model or harness. It’s: who builds the harness that never stops improving?
The harness is the product. The swarm is the moat.
Today, deep research harnesses are built in three ways — each with a structural ceiling:
- Big Model Labs (OpenAI DR, Gemini DR) tightly couple model and harness. They are locked to their own model, even when others excel at specific stages. And harness optimization competes for bandwidth with every other company priority.
- Centralized harness providers (Perplexity) decouple the harness from any one model, but every optimization flows through a single team’s bandwidth.
- Open-source projects (Perplexica) open development to the community — but without economic incentive, adaptation is voluntary and sporadic.
On a Bittensor subnet, all three ceilings break: the harness is decoupled from any model, development is open to anyone, and persistent economic incentives ensure miners adapt in real-time — because their income depends on it.
A competitive swarm with skin in the game will always adapt faster than a single team, a single release cycle, or an unpaid community can. The harness sets the bar. The swarm clears it — continuously.
We are building Harnyx (SN 67) that turns deep research into a commodity: an API that returns intelligence with provenance — faster and cheaper than any centralized alternative. Every miner’s local optimization compounds into global superiority, one query at a time.