NVIDIA gives AI agents verified skills for BlueField

NVIDIA has published a set of DOCA AI agent skills on GitHub, aimed at stopping coding agents from guessing at data processing unit work. The skills, released on 1 October 2026, hand agents verified API signatures, hardware capability requirements and build constraints for DOCA, the software platform behind NVIDIA's BlueField DPUs.

Why general-purpose agents stumble on DPUs

Agentic coding tools are built around general software habits, and DOCA is not general software. It spans accelerated networking, AI-native storage, in-silicon security, telemetry and lifecycle management, so an assistant without domain knowledge will happily invent an API call that cannot run on the hardware in front of it. Every one of those mistakes costs a compile, a flash or a debugging session.

The numbers NVIDIA is quoting

In a 65-prompt evaluation, NVIDIA says agents without the skills satisfied just 19 per cent of its checklist items, while agents running with the skills satisfied 100 per cent across every prompt. In a side-by-side build, the with-skills agent used 73 per cent less handwritten code and 46 per cent fewer hardware commands to get a Go-based RDMA application running on BlueField-3.

What is actually in the skill pack

The skills use a lightweight open format built around a SKILL.md file, and they cover the DOCA library including Flow, GPUNetIO, PCC and RDMA. That lets an agent check whether a given device supports a feature before writing code, and apply preflight checks before touching firmware-level settings. The repository sits under NVIDIA's own GitHub organisation and is powered by Nemotron.

Our opinion

The interesting number here is not 100 per cent, it is 19. NVIDIA is quietly admitting that a frontier coding agent left to its own devices is wrong about DOCA roughly four times out of five, which is a rough advertisement for where agentic development actually stands once you leave well-documented web frameworks. Shipping the fix as a public skill file rather than as a proprietary assistant is the smarter play, because it makes every editor and every model a potential DOCA developer. The catch is maintenance: hardware capability tables rot quickly, and a skill file that silently describes last year's BlueField is worse than no skill file at all. Whoever keeps SKILL.md current is doing more for adoption than any benchmark slide.