Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills

NVIDIA has made DOCA AI agent skills available on GitHub, describing them as a structured, verified foundation for AI agents working with its DOCA software platform for BlueField data processing units.
The skills use an open format built around a SKILL.md file containing real API signatures, hardware capability requirements and build constraints. NVIDIA says they span the full DOCA library, including Flow, GPUNetIO, PCC and RDMA. Each skill is scoped to a specific DOCA component or workflow, giving agents details such as function signatures, pkg-config module names, build-container constraints, and common failure modes with mitigations.
NVIDIA ran a 65-prompt evaluation of real DOCA developer tasks, graded against a required-answer checklist. Without the skills, agents satisfied only 19% of graded checklist items; with skills, they satisfied 100% across all prompts. Reported failure modes without skills included misuse of APIs and flags in 59 of 65 prompts, hardware capability not verified in 46, wrong tool routing in 39, skipped smoke tests in 34, and guessed versions in 30. NVIDIA also said the with-skills result was better across the 63 prompts testing API accuracy.
The post cites four scenarios, each illustrated with an evaluation prompt: using real DOCA Comch API calls, verifying device support before writing code for a GPUNetIO RDMA WRITE scenario, resolving a link-time undefined reference to doca_flow_init using pkg-config, and applying preflight checks, a rollback plan and cold power-cycle awareness before an mlxconfig-class firmware change on a live BlueField-3 DPU.
In a side-by-side demo, NVIDIA said two agents were given the same task of building a Go-based RDMA application on BlueField-3; the with-skills agent used 73% less handwritten code (189 lines versus 695) and 46% fewer hardware commands (20 versus 37).
DOCA is described as the unified software platform for BlueField DPUs, spanning accelerated networking, AI-native storage, in-silicon security, telemetry and lifecycle management.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
AI agents are becoming a standard part of development workflows, but general-purpose agents weren't built with specialized infrastructure software such as...