NVIDIA has published the clearest before-and-after this series has seen for agent skills on specialist software. In Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills, Claudia Martinez, Matan Raz and Tamir Tevet describe skills for DOCA, the software platform for BlueField data processing units, now on GitHub. Each is a SKILL.md scoped to one DOCA component, carrying “real API signatures, hardware capability requirements, and build constraints.” The diagnosis is the sentence worth keeping: without them, an agent works “from pattern-matching across general training data,” and “there is no machine-readable contract for agents to reason from.”
The figures are NVIDIA’s own evaluation, 65 real developer prompts graded against a required-answer checklist, and they should be read as a vendor testing its own skills. Without skills, agents misused APIs and flags on 59 of 65 prompts, skipped hardware capability checks on 46, routed to the wrong tool on 39, skipped smoke tests on 34 and guessed versions on 30. They satisfied 19 percent of checklist items without skills and 100 percent with them. What survives the marketing is the failure taxonomy, which matches what any team running agents against a niche SDK will recognise: invented functions, unverified hardware, code that compiles but will not link, and firmware changes made “without preflight checks, rollback plans, or knowledge of cold power-cycle requirements.” The skills move those checks before the code is written, not after it runs.
The catalog has the thinnest possible record to set against this. The Nvidia provider page carries one API page, the NVIDIA Developer API, and DOCA has no page of its own. One Agent Readiness dimension is lit, the rate limit signal. Agent skills, the dimension this post is entirely about, is unlit, because the skills live in a GitHub repository the record does not yet point at. The inference side of the company is held separately on the NVIDIA NIM provider page, which scores 68.0, exemplar band, with agent skills already lit.
The corporate record’s Kin Score is 11.8, emerging band, with discoverability at 64.3 and contract quality at 0.0, and Agent Readiness is 2.5, human-only. That is a pointer gap more than a company gap, and it is the post’s own argument turned around. NVIDIA writes that agents fail on DOCA because no machine-readable contract tells them what the hardware and software support. The catalog fails on DOCA for the same reason: nothing in the record tells it where the skills, the API reference or the capability manifests live. NVIDIA has given agents a contract to reason from. The record still needs one line telling an index where to find it.