NVIDIA DOCA Agent Skills Give AI Agents Verified APIs for BlueField Development

NVIDIA released DOCA AI agent skills on GitHub, a set of machine-readable specifications that give AI agents verified API signatures, hardware capability requirements, and build constraints for the DOCA software platform on BlueField DPUs. The skills cover the full DOCA library, including Flow, GPUNetIO, PCC, and RDMA, and are built around a SKILL.md file that agents can reason against directly.

In a 65-prompt evaluation run by the NVIDIA team, agents without skills satisfied only 19% of graded checklist items on real DOCA tasks, while agents with skills satisfied 100% across all prompts. The most common failure mode without skills was API misuse, affecting 59 of 65 prompts, followed by unverified hardware capability in 46 prompts, wrong tool routing in 39, skipped smoke tests in 34, and guessed versions in 30.

What the skills change

Each skill is scoped to a specific DOCA component or workflow. When an agent loads a skill for DOCA Flow, it gets real function signatures, correct pkg-config module names, build-container constraints, and common failure modes with mitigations. This replaces pattern-matching against general training data, which lacks the hardware-specific contracts of the DOCA library.

The skills address four development scenarios. For API accuracy, the with-skills agent uses only real DOCA Comch API calls with correct argument order and flags, avoiding invented functions that fail at runtime. For hardware verification, the agent checks GPU-NIC PCIe topology and confirms GPUNetIO support before writing code, preventing mismatches discovered only on real hardware.

For build correctness, the agent uses pkg-config to look up correct linker flags for doca-flow, resolving undefined references like doca_flow_init without sending developers down wrong paths. For firmware-level changes, the agent applies preflight inventory, assumes an out-of-band path as a precondition, requires an explicit maintenance window, includes a rollback plan, and notes that mlxconfig-class writes take effect only on a cold power cycle, not a warm reboot.

Measured impact

A side-by-side demo tasked two agents with 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. The without-skills agent met none of the requirements for the firmware-level scenario, while the with-skills agent met every requirement.

The skills are available in the NVIDIA/skills GitHub repository, with general skills for a quick onramp and library-specific skills for DOCA Flow, GPUNetIO, RDMA, and more. For teams using AI agents on DOCA applications at scale, the reduction in correction cycles compounds across every developer, task, and deployment.