NVIDIA Releases DOCA Agent Skills to Improve BlueField Infrastructure Development

NVIDIA has released a set of DOCA AI agent skills on GitHub designed to help coding agents build applications for NVIDIA BlueField data processing units with fewer unsupported API calls and hardware-related errors. The move targets a narrow but consequential problem: general-purpose AI coding agents often lack the device-specific knowledge required for infrastructure software.

The skills package gives agents structured information about APIs, hardware capabilities, build requirements and deployment constraints across the NVIDIA DOCA software platform. NVIDIA says the approach can reduce correction cycles when developers build networking, storage, security and telemetry applications for BlueField systems.

What changed for BlueField developers

NVIDIA DOCA is the company’s software platform for developing applications on BlueField DPUs. It covers functions including accelerated networking, AI-oriented storage, in-silicon security, telemetry and lifecycle management. The new DOCA AI agent skills are intended to give coding agents a more dependable technical foundation than general training data alone.

Each skill is delivered in a lightweight, open format centered on a SKILL.md file. According to NVIDIA, the files contain verified API signatures, hardware capability requirements, build-container constraints, package names and known failure modes. The skills are scoped to particular DOCA components or workflows rather than presented as a single broad instruction set.

For example, a skill for DOCA Flow can provide the correct function signatures and pkg-config module names, while also telling an agent which build conditions and common mitigations apply. NVIDIA says the skills cover the broader DOCA library, including DOCA Flow, GPUNetIO, PCC and RDMA.

The files do not replace the coding agent itself. Instead, NVIDIA describes them as a machine-readable layer of domain knowledge that lets an agent check its proposed code against the relevant software and hardware constraints before producing or deploying it.

NVIDIA’s evaluation shows a large gap—but it is vendor-reported

NVIDIA tested agents on 65 DOCA development prompts, ranging from short questions to multi-requirement implementation tasks. The company graded responses against task-specific pass/fail checklists.

In NVIDIA’s evaluation, agents without the skills satisfied 19% of the checklist items, while agents using the skills satisfied 100% across all 65 prompts. The company said unsupported functions, incorrect flags and invalid image tags were recurring problems in the unassisted responses. It reported that API misuse appeared in 59 of the 65 prompts.

Those figures indicate the potential value of supplying agents with authoritative, structured context for specialized infrastructure work. They should also be read as NVIDIA’s own benchmark, not as an independent assessment of every coding agent or DOCA workflow. The source material does not identify all of the models, agent configurations, scoring details or external reviewers involved in the test, so the results cannot yet establish how the skills will perform across different tools and production environments.

NVIDIA also presented a side-by-side demonstration involving a Go-based RDMA application on a BlueField-3 system. The company said the agent using the skills required 73% less handwritten code and 46% fewer hardware commands than the agent without them. These are likewise vendor-reported results from a demonstration, rather than a general productivity study.

Why the details matter to infrastructure teams

The problem NVIDIA is addressing is more serious in infrastructure development than in many application-level coding tasks. An invented function may produce a visible compilation error in a conventional software project. On a DPU, an incorrect assumption can also involve device support, firmware state, container compatibility, link configuration or a change that requires a restart or power cycle.

NVIDIA says the skills allow agents to verify device support before writing code and apply preflight checks before making firmware-level changes. They can also guide rollback planning and account for situations in which a cold power cycle is required. Those checks are relevant to teams that cannot treat physical infrastructure as an infinitely repeatable development environment.

For developers, the practical benefit is not simply fewer lines of generated code. The larger potential gain is reducing the number of cycles between an agent’s first answer and a working application on real hardware. A networking engineer building a DOCA Flow pipeline or an RDMA developer configuring a host-to-DPU workflow may spend less time correcting invented interfaces and more time validating behavior.

For enterprise buyers, however, the skills do not remove the need for testing, access controls or human review. A machine-readable description of an API can improve an agent’s starting point, but it cannot guarantee that generated code is safe for a production network, compatible with a particular firmware release or appropriate for a deployment’s security policy.

Open skills as a distribution strategy

NVIDIA is making the DOCA skills available through the NVIDIA/skills GitHub repository, which gives developers a direct way to inspect and use the material alongside their preferred AI coding tools. The repository-based model also creates a path for the guidance to evolve as DOCA interfaces, supported hardware and build environments change.

That distribution choice matters because specialized infrastructure agents are only useful if their instructions stay aligned with the software they operate. NVIDIA’s description of the skills emphasizes verified contracts and hardware requirements rather than broad conversational expertise. Maintaining that accuracy will be essential as BlueField platforms and DOCA components develop.

The announcement also reflects a broader direction in AI-assisted development: instead of expecting a general model to memorize every specialized system, vendors are packaging domain context for agents to load at task time. In this case, the domain is not a consumer application framework but a hardware-software stack where incorrect assumptions can delay deployment.

What to watch next

The first signal will be whether developers outside NVIDIA reproduce the reported improvements across different coding agents, models and DOCA versions. Independent evaluations should clarify whether the 19% versus 100% checklist gap holds beyond NVIDIA’s prompt set.

Teams should also watch how quickly the repository tracks changes to BlueField hardware, firmware, container images and API behavior. Stale skills could reintroduce the same reliability problems they are intended to reduce.

A third signal is adoption in complete workflows rather than demonstrations. Evidence that skills help with testing, debugging, rollback and production deployment—not only initial code generation—would show whether they reduce operational risk as well as developer effort.

Creati.ai perspective

NVIDIA’s release is a targeted response to a real limitation of general-purpose coding agents: fluency is not the same as accurate knowledge of a specialized hardware platform. By exposing API and device constraints in a structured format, NVIDIA is attempting to move agent assistance closer to an engineering tool and farther from plausible code generation.

The strongest claims remain vendor-controlled, and the value of the skills will depend on maintenance, independent testing and integration with deployment safeguards. Still, the approach offers a practical template for infrastructure vendors: provide agents with versioned, machine-readable operating knowledge instead of relying on models to infer critical details from generic training data.