When models jump in capability, they internalize the scaffolding built to compensate for their weaknesses — RAG hacks, CoT, output-parsing retries, multi-agent frameworks ("wrapper death"). But this only eats one type.
- Prosthetic scaffolding compensates for a capability gap. Anything with a
token → tokensignature is internalizable (Sutton's bitter lesson). Gets eaten. - Constitutive scaffolding provides what is outside capability — authority bits, external state handles, trust boundaries, verification, audit. A stochastic substrate cannot self-certify (promise ≠ proof; the verifier must be external to the verified). Not eaten.
The type-theoretic discriminant: does it need an authority bit or an external-state handle? If yes, it's not in the token→token signature, so capability growth can't absorb it. "AI didn't escape Wirth's Law, it just installed a garbage collector for one type of complexity."
This is the kernel/userland split of harness-is-the-os-of-the-intent-stack.
Logan Kilpatrick's "harness dies in 12 months" is exactly the prosthetic prediction (2026)
Logan Kilpatrick (Google DeepMind), in a Sequoia interview (recapped 2026-07, youtu.be/EciNo2uhWRk), gave the mainstream version of the prosthetic claim: "the model will eat the scaffolding — the independent value of the Agent Harness drops sharply within 12 months," because the model natively absorbs tool-calling, code execution, search, hosted services. Every capability he names as absorbed is token → token / prosthetic — so by this note's discriminant he is right about that layer. It's not a refutation of the split; it's its clearest public statement.
His own supporting points land on the constitutive side (i.e. they reinforce what does not get eaten):
- "a truly general model should adapt to different harnesses (else it isn't general)" → the harness survives as an interface / authority contract the model adapts to; it isn't erased.
- his call for a "Harness benchmark" → constitutive harness gets standardized and evaluated, not absorbed.
- "jagged super-intelligence: verifiable domains break through first" → verification (constitutive — verifier external to verified) is the frontier, not a crutch.
- autonomous agents "handling billions of tokens, refactoring an OS" → more authority/audit/recovery is needed as autonomy rises, not less.
Caveat on the word: "harness" is two-typed. Logan's dying harness = the prosthetic plumbing; the constitutive harness (authority / verification / audit — the recursive-harness control layer) is the durable moat his own closing question ("which capabilities are the true durable moats?") points straight at. Don't let one word collapse the two.