Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills

NVIDIA Developer blog introduces DOCA AI agent skills on GitHub: machine-readable SKILL.md packages with verified API signatures, hardware capability requirements, and build constraints across Flow, GPUNetIO, PCC, RDMA, and more. In a 65-prompt internal eval, agents without skills hit 19% of checklist items; with skills, 100%.

Common no-skill failures included API/flag misuse (59/65), unverified hardware capability (46/65), and wrong tool routing. A side-by-side RDMA-on-BlueField-3 demo claimed 73% less handwritten code and 46% fewer hardware commands with skills.

This is useful AI done the unsexy way — give the agent a contract for a gnarly domain instead of hoping Stack Overflow vibes survive firmware writes. We’re strongly for skills/packaged expertise when the alternative is inventing doca_flow flags on a production DPU.

The 100% number is NVIDIA grading NVIDIA. Still, the failure taxonomy is believable: infrastructure libraries move fast, and general agents hallucinate APIs. Shipping skills for mlxconfig-class changes with preflight, OOB path, and cold power-cycle awareness is exactly where agent coding should grow up.

I’m watching whether the GitHub skills stay current as DOCA versions churn — stale skills are confident wrongness. If they do, this pattern (domain skill packs for agents) will matter far beyond BlueField.

DOCA is NVIDIA’s software platform for BlueField DPUs spanning networking, storage, security, and telemetry — now a target for agentic coding workflows.