NVIDIA Introduces DOCA Agent Skills to Supercharge Infrastructure AI
NVIDIA has released DOCA AI agent skills on GitHub to give development AI agents specialized domain knowledge for BlueField data processing units, drastically cutting down errors.
As artificial intelligence agents become increasingly common in software engineering, standard general-purpose models often struggle with specialized infrastructure platforms. Without domain-specific context, these agents frequently resort to guesswork, leading to frustrating correction cycles that slow down deployment. To bridge this gap, NVIDIA has released DOCA AI agent skills on GitHub, designed specifically to provide a structured, verified foundation for infrastructure development on NVIDIA BlueField data processing units.
DOCA serves as the unified software platform for BlueField DPUs, covering accelerated networking, in-silicon security, AI-native storage, telemetry, and lifecycle management. However, because the DOCA software library is extensive and hardware-specific, standard AI models lacking specialized training frequently misuse APIs, skip hardware verification, guess version numbers, and neglect smoke tests. NVIDIA evaluated general-purpose agents across 65 real development prompts and found they successfully met only 19% of graded checklist items without additional domain guidance.
Equipping AI Agents with Domain Knowledge
The newly available DOCA AI agent skills solve these shortcomings through a lightweight, open format built around a SKILL.md file. This file contains real API signatures, hardware capability requirements, and build constraints for various components such as Flow, GPUNetIO, and PCC. Rather than replacing the agent, these skills supply a machine-readable operational framework that lets the model reason like an experienced developer.
When NVIDIA re-tested the same 65 prompts using agents equipped with DOCA skills, the performance reached a 100% success rate across all evaluated tasks. In side-by-side comparisons, an agent with DOCA skills produced programs using 73% less handwritten code and roughly half the hardware commands compared to an unequipped agent.
What it means for developers
For engineers working with complex networking and DPU hardware, these agent skills offer practical benefits by reducing trial-and-error debugging. Developers can build faster because agents receive verified API calls from the start, avoiding non-existent functions or incorrect flags. Furthermore, developers can try top AI models cheaply through one API at https://apixoai.online.
Additionally, the skills ensure agents check hardware capabilities before writing code, preventing mismatches between software assumptions and physical device limits. They also help agents properly manage build-time requirements, like correct pkg-config linker flags, and execute safe procedures for high-risk operations such as firmware-level parameter updates on live hardware.
Getting Started with DOCA Skills
By incorporating preflight checks, rollback plans, and strict adherence to build constraints, DOCA agent skills turn risky infrastructure tasks into disciplined procedures. Development teams looking to accelerate their workflows can access the new framework through the official NVIDIA/skills GitHub repository, where both general onramp materials and library-specific configurations are now available.
Source: Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills — NVIDIA Technical Blog. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
