Humans own intent and judgment

The human’s remaining work is surfacing the choices and non-functional requirements that define good work locally, and delivering them in a form the model can learn in context. A clearly specified task is the one thing models reliably solve; under-specification is the named cause of disappointing agentic outcomes. Empowering an agent is like onboarding a capable colleague who does not yet know what good looks like here.

When a model returns thirty plausible answers to one research question, each worth months of work, the bottleneck moves to deciding which deserves attention. Taste resists automation hardest — novelty, elegance, longevity are exactly the qualities a good reviewer judges and nobody has automated. Hiring has already shifted away from coding tests toward problem decomposition, prompting skill, and the ability to detect confidently-wrong output.

Reads are free, writes are gated, somebody clicks approve — with identity attached to the click. Reviewing the agent’s intermediate reasoning is waste; reviewing its plan before it executes is not. The efficiency case here is not only governance: a dedicated plan-review phase converged to the same quality in roughly 35 iterations instead of 175, one of the largest measured token savings in either scan.

The brief is the specification artifact; its Verify rows are the evaluation authored before the work; the human-gated merge is the act-step gate; and the desk roles never gate the agent’s intermediate thinking. Ema’s and SUSE’s articulations are crisper than Assay’s own docs, and identity-stamped approvals in infrastructure operations are outside confirmation that this governance shape is converging across domains — the same design arrived at independently. The 175-to-35 result is the first quantified argument for authoring briefs at all: planning before executing is not process overhead, it is a 5x efficiency measure.

Intake has no readiness test for specification quality — a check that a brief unable to state its non-functional requirements is not yet dispatchable is designed and unbuilt, and it is exactly what the harness-engineering evidence describes. Waves are static once authored; the verifier-feeds-replanner structure suggests verification outcomes should be able to split or merge pending briefs, which is an extension of the dreaming pass into plan maintenance rather than a new mechanism. And the sycophancy point cuts at the review loop: the reviewing App’s verdict is only worth something if it forms independently of the worker’s self-report — which is the design, but is not measured.

The capital-markets panel raises the open question worth carrying: judgment is traditionally built by making your own mistakes, so where does the next generation’s come from if the machine does the experimenting? Read against Assay’s lifecycle, that argues the human merge gate should be staffed deliberately as the place judgment is acquired, not treated as pure cost to be minimized.

• Whether anyone reproduces the plan-review efficiency result outside NVIDIA’s own stack; a second 5x-class number would make plan review a required stage rather than a good practice.
• Whether a specification-readiness gate appears in any harness — a check that refuses to dispatch an under-specified task, rather than a linter on the resulting code.
• Whether the eval frontier’s move toward deliberately under-specified tasks undercuts the specify-first doctrine, or simply measures a different capability.