Thread: Why do independent vendors converge on “harnesses thin as models improve” — and what does thinning actually leave behind?
The question
A same-day 7-video batch (2026-07-15 ingest) landed three sources that each touch the agent-harness concept’s existing “subtraction principle” material from a different angle: Jensen Huang (NVIDIA), interviewed by Harrison Chase (LangChain) frames companies as built “on harnesses, not business processes” and describes post-training a model inside a harness; Anthropic’s own Claude Platform team gives the wiki’s first first-party interview account of harnesses getting thinner as models improve, and names “meta-harnesses” as what teams build once individual harnesses thin; Ishit Vachhrajani (AWS) doesn’t use “thinning” vocabulary directly but describes Amazon Bedrock AgentCore as a trust-infrastructure layer that hardens even as the “advanced team structures” franchise’s own economics (USE/COMPOSE/BUILD) assume the business-logic layer keeps shrinking.
This lands on top of an existing, unresolved wiki debate: agent-harness’s Debates section already holds a live tension between Kilpatrick (Google DeepMind, June 2026) — who predicts the model absorbs the harness natively within ~12 months, so harness-alpha is transient — and Chatterjee’s “model rented, harness owned” framing, where the harness is a durable, compounding asset. Both agree on the mechanism (harness components encode assumptions about model limitations, and those assumptions expire); they disagree on the strategic conclusion (does the harness eventually vanish, or does it persist in a different shape?).
The question this thread tracks: Do the three new July 2026 sources settle, sharpen, or merely restate the Kilpatrick/Chatterjee tension — and specifically, when a harness thins, what (if anything) is left, and does it thin uniformly or by layer?
Why this thread opened-and-closed in the same session
The three sources, plus the wiki’s existing agent-harness material (subtraction principle, the Prompt Engineering empirical anchor, the Kilpatrick debate), were fully read and cross-referenced during the same-day ingest that landed them — see the 7-video batch processed 2026-07-15. No further sourcing was needed to answer the question at hand; per CLAUDE.md §Synthesis, this thread is the schema-compliant antecedent to the synthesis below rather than a substantive open investigation.
Closed
Filed harness-thinning-what-persists synthesis on 2026-07-15. Headline finding: thinning is layer-specific, not uniform — the business-logic content of a harness thins as models improve, while the identity/trust/access-control substrate around it hardens. This resolves the Kilpatrick/Chatterjee tension partially: Kilpatrick is right about the business-logic layer (it does erode toward the model); Chatterjee is right about the governance layer (it does not — if anything, three independent vendors are actively building more of it, not less). Neither “the harness vanishes” nor “the harness is the permanent moat” is correct as a single claim about the whole harness; both are correct about different halves of it.