NVIDIA Launches DOCA Agent Skills for BlueField DPUs

NVIDIA has launched DOCA AI agent skills, a set of machine-readable specifications designed to help AI agents write code for its BlueField data processing units (DPUs). Available on GitHub and powered by the NVIDIA Nemotron model, these lightweight skills use a standardized markdown format to supply general-purpose AI agents with verified API signatures, hardware requirements, and build constraints. This domain-specific knowledge covers key DOCA libraries such as Flow, GPUNetIO, PCC, and RDMA, preventing agents from relying on guesswork when developing infrastructure software.

To measure the impact of these skills, NVIDIA evaluated AI agents across 65 developer prompts. Without the specialized skills, agents satisfied a mere 19 percent of the required checklist items, frequently inventing nonexistent APIs in 59 of the prompts and failing to verify hardware compatibility in 46 instances. In contrast, agents equipped with the new DOCA skills achieved a perfect 100 percent success rate across all 65 test prompts. Build correctness was also tested in 10 of these prompts, where the skilled agents successfully navigated complex compilation and linking constraints.

For developers, this translates to a massive reduction in debugging and manual coding. In a side-by-side demonstration building a Go-based RDMA application on a BlueField-3 DPU, an agent utilizing the skills required 73 percent less handwritten code, generating 189 lines compared to the 695 lines written by the unskilled agent. Additionally, the skilled agent executed 46 percent fewer hardware commands, dropping from 37 to 20.

Beyond writing cleaner code, the skills allow agents to perform preflight checks and safely execute firmware-level changes. When tasked with writing mlxconfig-class parameters on live hardware, the skilled agent successfully accounted for out-of-band paths, maintenance windows, rollback plans, and cold power-cycle requirements. This structured approach ensures that developers can deploy stable infrastructure code with fewer correction cycles.