NVIDIA 发布 DOCA AI agent skills，让智能体在 BlueField 上开发更高效

Technical Blog

sessionStorage.removeItem('facets.search.text'); sessionStorage.removeItem('facets.search.text-copy'); sessionStorage.removeItem('facets.search.filters');

NVIDIA DOCA AI agent skills provide verified API signatures, hardware capability requirements, and build constraints so agents can reason like experienced DOCA developers. In a 65-prompt evaluation, agents without skills satisfied only 19% of checklist items, while agents with skills satisfied 100% across all prompts. Skills cover the full DOCA library including Flow, GPUNetIO, PCC, and RDMA, enabling agents to verify device support before writing code and apply preflight checks for firmware-level changes. A side-by-side demo showed the with-skills agent used 73% less handwritten code and 46% fewer hardware commands to build a Go-based RDMA application on BlueField-3. Next Steps Visit the NVIDIA/skills GitHub repository to access DOCA AI agent skills for Flow, GPUNetIO, RDMA, and more.

Powered by NVIDIA Nemotron. AI-generated content may summarize information incompletely. Verify important information. Learn more

AI agents are becoming a standard part of development workflows, but general-purpose agents weren’t built with specialized infrastructure software such as NVIDIA DOCA in mind. Without domain-specific knowledge, agents may fall back on guesswork. This is an issue in infrastructure development because every correction cycle takes time away from deployment.

DOCA is the unified software platform that unlocks the full potential of NVIDIA BlueField data processing units (DPUs) for agentic AI infrastructure. It spans accelerated networking, AI-native storage, in-silicon security , telemetry, and lifecycle management. It’s the development platform for teams building on NVIDIA BlueField.

DOCA AI agent skills are now available on GitHub. These skills provide a structured, verified foundation to address where general-purpose AI agents fall short on DOCA development. They provide agents with verified API signatures, hardware capability requirements, and build constraints.

DOCA AI agent skills are a standardized way to give AI agents new capabilities and expertise. The lightweight, open format is built around a SKILL.md file containing real API signatures, hardware capability requirements, and build constraints. Together, the skills provide agents with a reasoning and operational framework for DOCA.

Skills span the full DOCA library, including Flow, GPUNetIO, PCC, and more. They don’t replace the agent, but give it the domain knowledge to reason like an experienced DOCA developer.

Each skill is scoped to a specific DOCA component or workflow. When an agent loads a skill for DOCA Flow, for example, it gets the real function signatures, the correct pkg-config module names, the build-container constraints, and the common failure modes and their mitigations. The skills are not just a summary of documentation, but a machine-readable specification the agent can reason against directly.

When you ask a general-purpose AI agent to set up a DOCA Comch (Comm Channel), configure an RDMA context, or debug a link failure in a DOCA Flow program, the agent is working from pattern-matching across general training data—not from verified DOCA API contracts, hardware capability manifests, or build system specifications. The DOCA library surface is large, rapidly evolving, and hardware-specific in ways that general training data does not capture. There is no machine-readable contract for agents to reason from, no guaranteed stable interface between what the agent knows and what the hardware and software actually support.

To quantify the gap, the NVIDIA team ran 65 real DOCA developer prompts comparing agent performance with and without the skills. These prompts ranged from a one-line question to detailed multi-requirement prompts, graded against a required-answer checklist (a specific set of pass/fail criteria for each task). Without the skills, agents consistently made the same mistakes, including:

These are consistent failures. Without skills, agents satisfied only 19% of graded checklist items on real DOCA tasks. With skills, they satisfied 100% across all 65 prompts. Without the skills, every agent mistake becomes a developer debugging session, and a longer path to working code.

DOCA agent skills allow you to equip agents with domain expertise. This means you can build faster, ship more stable code, and deploy with fewer unknowns.

The combined effect is fewer correction cycles between task and working code. For teams using AI agents on DOCA applications at scale, that reduction compounds across every developer, task, and deployment.

Each of four capabilities are detailed in the following sections. Each capability is illustrated with a real prompt from our 65-prompt evaluation and the benefits of using the skill.

What DOCA AI agent skills address: Misuse of APIs: agents inventing functions, flags, and image tags that don’t exist.

This was the most frequent failure mode in our evaluation, affecting 59 of 65 prompts. Without skills, agents generate code referencing DOCA functions that don’t exist in the library, flags with incorrect names, and image tags that fail at runtime. As the developer, you end up debugging an error you didn’t introduce (the agent’s invention).

Prompt: “I have a host with a BlueField-3 DPU and DOCA installed. I want to set up a DOCA Comch (Comm channel) between a host-side process and a DPU-side agent so they can exchange small control messages (under 4 KiB each, a few times per second).”

The AI agent with skills uses only real DOCA Comch API calls: correct argument order, verified flags and lifecycle sequence, nothing invented.

When the agent’s first response uses the correct API surface, you can move directly to integration rather than starting with a debugging session. Across 63 prompts that tested API accuracy, the with-skills result was better every time.

What DOCA AI agent skills address: Agents writing code for features the device doesn’t support.

Hardware capability verification was missed in 46 of 65 prompts without skills. The failure pattern is consistent: the agent assumes a capability is present, generates code against it, and the developer discovers the mismatch only when running on real hardware. At that point, the fix requires understanding which capability check was skipped, reworking the code path, and retesting.

Prompt: “I have two hosts wired together with a high-speed InfiniBand fabric. Each host has one NVIDIA GPU and one ConnectX NIC, and DOCA is installed on both. I want to measure the latency of an RDMA WRITE work request when the WR (work request) is posted from a CUDA kernel on the GPU (not from a host CPU thread) so I can decide whether to commit my application to the GPUNetIO Verbs interface.”

The AI agent with skills checks what the device actually supports before writing any code. It verifies GPU-NIC PCIe topology, confirms GPUNetIO support before committing, and names alternative API surfaces and when each applies.

The AI agent with skills validates the preconditions for the answer against the actual hardware in use, before any code is written.

What DOCA AI agent skills address: Agents producing code that won’t compile or link in the DOCA container.

Build correctness was tested on 10 of 65 prompts. This is a smaller subset, but when it applies, development cannot proceed.

Prompt: “I’m trying to build a DOCA Flow program from one of the shipped samples. The compile step works but the link fails: undefined reference to doca_flow_init. How do I debug this?”

The AI agent with skills identifies the undefined reference as a link-time failure, uses the build tool (pkg-config) to look up the correct linker flags for doca-flow.

An agent that generates an incorrect linker flag sends the developer down a path that won’t resolve. The with-skills agent immediately names the correct diagnostic approach and the right tool to retrieve the correct flags.

What DOCA AI agent skills address: Agents applying firmware-level changes on live hardware without preflight checks, rollback plans, or knowledge of cold power-cycle requirements.

The more complex the task, the more the skills matter. In our 65-prompt evaluation, the largest performance gap appeared in the highest-complexity scenario: firmware-level changes on live hardware.

Prompt: “I have a production BlueField-3 DPU running a DOCA workload. I loaded the DOCA AI agent skill for my service, and the next step is to write an mlxconfig-class firmware-level parameter.”

The AI agent with skills applies the full discipline required for firmware-level changes to meet every requirement. This includes preflight inventory, assuming an out-of-band (OOB) path as a precondition, an explicit maintenance window, a rollback plan, and notes that mlxconfig-class writes take effect only on a cold power cycle, not warm reboot.

The with-skills agent met every requirement. The without-skills agent met none. On a live production DPU, none of these misses are recoverable debugging steps. The skills carry the knowledge of what can go wrong and why, turning a risky firmware change into a disciplined, safe procedure.

Across all 65 prompts, AI agents with DOCA skills reached the correct answer every time. This means fewer correction cycles and faster time to working code.

The demo in Video 1 puts the efficiency gain in concrete terms. Two agents were given the same task: build a program using NVIDIA DOCA to send real RDMA traffic on BlueField-3. Both succeeded, but the with-skills agent used 73% less handwritten code (189 lines versus 695) and roughly half the hardware commands (20 versus 37, a 46% reduction). That’s the difference between an agent that rediscovers DOCA APIs and build requirements by trial and error, and one that already has them before writing a single line.