The First and Last Harness Is Us
What it means
Zooming out: three ways to read the belt
So far this is an explainer. But the harness matters far beyond product design, and three very different thinkers help show why.
The harness as a growth engine. In Situational Awareness, Leopold Aschenbrenner argues that AI progress comes from three sources: more compute, better algorithms, and what he calls unhobbling. Unhobbling is the harness: tools, scaffolding, step-by-step reasoning, using a computer, getting onboarded to a job. His bet is that these gains turn a chatbot into a drop-in remote worker, and that they're as big a deal as the model itself.
The harness as the last place difference lives. A 2025 paper by Turegeldinova and colleagues describes AI as a centripetal technology, one that pulls things toward the center. When everyone builds on the same handful of models, products start to converge on the same template. If that's right, the model stops being where you compete. The harness is what's left: your tools, your data, your workflow, your taste.
The harness as a question of ownership. Balaji Srinivasan's The Network State is about who gets to leave. Apply it here and the question is blunt: whoever owns the harness owns your agent's memory, its connections, and its permissions. If you can pick that up and move it to another model, you're a citizen. If you can't, you're a tenant.
Put together: All three are really theories about which technologies concentrate power and which spread it out. The open question of this decade is whether AI concentrates it. The harness is where that gets decided.
Forecast: what happens to the harness?
The obvious prediction is that smarter models won't need harnesses. I think that's half right, and the wrong half is the important one.
First, the scaffolding gets absorbed. Things harnesses do today to make up for weak models, like breaking a task into steps, choosing which tool to use, or checking their own work, will move inside the model. It already happened once: step-by-step reasoning started as a trick people wrote into prompts and is now built into the models themselves.
Second, the guardrails get heavier. As agents start acting on real accounts with real money, the harness grows everything a company grows around a powerful employee: identity, payment limits, approvals, audit trails, liability. The more capable the agent, the more of this it needs.
Third, the harness becomes the product. As base models converge, the companies that win are the ones with the best desks, not the best geniuses.
Fourth, the harness becomes your identity. Expect a real fight over who owns an agent's memory and whether you can take it with you. Open, portable harnesses against closed platforms will be the browser wars of this era.
Fifth, at the far end, the harness becomes containment. Aschenbrenner's plan for superintelligence includes locked-down data centers, AI systems monitoring other AI systems, and multiple sign-offs before anything big happens. That's a harness too, just pointed the other way. It doesn't help a weak mind do more. It keeps a strong mind from doing too much.
"we need to be in a position where alignment can fail—but failure wouldn't be catastrophic." — Leopold Aschenbrenner, Situational Awareness