NVIDIA launches DOCA Agent Skills with 100% task compliance for hardware-aware AI coding agents
NVIDIA has launched DOCA Agent Skills, a collection of specialized capabilities that allow AI coding agents to work directly with NVIDIA BlueField infrastructure using verified, hardware-aware knowledge rather than relying only on general-purpose model training.
The skills package gives coding agents structured access to NVIDIA DOCA documentation, APIs, development patterns, hardware capabilities and implementation guidance. It is designed for agents building applications that run on BlueField DPUs and related NVIDIA infrastructure.
In NVIDIA's evaluation across 65 development prompts, agents without DOCA Agent Skills satisfied only 19% of the required checklist items. With the skills enabled, NVIDIA reports that the agents completed 100%.
In one development workflow, NVIDIA also measured a 73% reduction in manually written code, indicating that the skills can affect both correctness and the amount of implementation work left to the developer.
··········
NVIDIA DOCA AGENT SKILLS AT A GLANCE
........
Specification | DOCA Agent Skills |
Developer | NVIDIA |
Product type | AI coding-agent skills |
Primary platform | NVIDIA BlueField |
Software ecosystem | NVIDIA DOCA |
Primary users | Developers and coding agents |
Core purpose | Hardware-aware software development |
Knowledge source | NVIDIA DOCA technical resources |
Evaluation size | 65 prompts |
Baseline checklist satisfaction | 19% |
With DOCA Agent Skills | 100% |
Improvement | +81 percentage points |
Manual coding reduction | Up to 73% in NVIDIA's example |
Compatible workflow | Agentic coding |
Distribution model | Reusable agent skills |
........
The results are NVIDIA-reported measurements, not independent benchmarks. They measure compliance with predefined development checklists rather than general software-engineering intelligence.
The gap nevertheless illustrates the problem NVIDIA is attempting to solve: even capable coding models can lack sufficiently precise knowledge of specialized hardware APIs and development environments.
··········
GENERAL CODING MODELS DO NOT AUTOMATICALLY KNOW SPECIALIZED HARDWARE
Modern coding agents can generate applications across common programming languages and frameworks because enormous amounts of relevant material were represented in their training data.
Specialized infrastructure creates a different problem.
BlueField development involves DOCA libraries, hardware accelerators, networking functions, security capabilities and platform-specific APIs that may change faster than a foundation model's training cycle.
A model can therefore generate syntactically plausible code while using an outdated API, selecting the wrong hardware capability or implementing a function inefficiently.
DOCA Agent Skills provide the agent with specialized instructions and reference material at inference time.
Instead of expecting the underlying model to have memorized the complete BlueField software stack, the agent can consult the skill while performing the task.
··········
THE SKILLS PACKAGE KNOWS HOW BLUEFIELD HARDWARE IS SUPPOSED TO BE USED
DOCA is NVIDIA's software framework for programming BlueField data processing units.
BlueField DPUs can offload infrastructure workloads from host CPUs, including networking, storage, security and data-center management functions.
That means correct software design depends on understanding more than the programming language. The agent needs to know which operations can be accelerated by the DPU, which DOCA APIs expose those capabilities and how the application should be structured around the underlying hardware.
DOCA Agent Skills supply that additional context.
The objective is therefore not merely better code completion. NVIDIA is attempting to give coding agents a machine-readable representation of how its infrastructure should actually be programmed.
··········
NVIDIA TESTED THE SKILLS ACROSS 65 DEVELOPMENT PROMPTS
NVIDIA evaluated the approach using 65 prompts representing DOCA development tasks.
Each task contained a checklist of requirements that the generated implementation was expected to satisfy.
Without the specialized skills, coding agents completed 19% of the checklist items.
With DOCA Agent Skills available, NVIDIA reports 100% checklist-item satisfaction.
........
NVIDIA evaluation | Without skills | With DOCA Agent Skills |
Development prompts | 65 | 65 |
Checklist satisfaction | 19% | 100% |
Percentage-point change | — | +81 pp |
Relative checklist completion | 1.0× | ~5.26× |
........
The ~5.26× figure is a Data Studios calculation, obtained by dividing 100 by 19.
It should not be interpreted as a 5.26× improvement in general coding performance. The benchmark measures whether the agent satisfies the checklist requirements defined for these particular DOCA tasks.
The result instead quantifies how much specialized platform knowledge changed performance inside NVIDIA's own evaluation.
··········
ONE WORKFLOW REQUIRED 73% LESS HAND-WRITTEN CODE
NVIDIA also provides a more practical development example.
In that workflow, using an AI coding agent equipped with DOCA Agent Skills reduced the amount of code that developers had to write manually by 73%.
That does not mean every BlueField project will require 73% less human programming.
The result depends on the task, the coding agent, the quality of the generated implementation and the amount of review or modification required afterward.
It does show the intended workflow: the developer describes the required infrastructure behavior, while the agent translates a larger portion of that specification into platform-specific implementation.
Human work moves toward architecture, validation, debugging and approval, while more of the repetitive implementation can be delegated to the agent.
··········
AGENT SKILLS ARE DIFFERENT FROM RETRAINING THE UNDERLYING MODEL
One of the important architectural characteristics is that DOCA Agent Skills do not require NVIDIA to train a new foundation model every time the platform changes.
The specialized knowledge exists as a layer that the agent can use during its workflow.
This creates a different update cycle.
When an API changes or a new BlueField capability becomes available, the relevant skill can be updated without waiting for the next generation of the coding model.
The same underlying agent can therefore become more capable in a specialized domain by receiving a better skill package.
This approach also avoids embedding every hardware platform and software library directly into model weights.
··········
SKILLS CAN REDUCE API HALLUCINATIONS BUT CANNOT ELIMINATE THEM
Hardware-aware context addresses one common weakness of AI-generated code: plausible but nonexistent interfaces.
A general model may recognize the structure of an API and generate a function name or parameter combination that looks reasonable but is not supported by the current software stack.
Giving the agent authoritative platform instructions reduces the need to reconstruct those details from model memory.
It does not make generated code automatically correct.
The agent can still misinterpret documentation, combine individually valid APIs incorrectly, misunderstand the developer's requirement or produce code with security and performance problems.
Compilation, testing, static analysis, runtime validation and human review therefore remain necessary for production systems.
··········
HARDWARE-AWARE AGENTS CAN OPTIMIZE FOR MORE THAN FUNCTIONAL CORRECTNESS
Infrastructure programming introduces another distinction from conventional code generation.
Two implementations can produce the same functional result while using the underlying hardware very differently.
A coding agent that understands BlueField can potentially select architectures that offload appropriate workloads to the DPU instead of leaving them on the host CPU.
That can affect CPU utilization, network processing, security isolation and overall system efficiency.
The value of hardware-aware skills therefore extends beyond generating code that compiles.
The longer-term objective is to allow agents to reason about the relationship between software architecture and the physical infrastructure executing it.
··········
THE APPROACH COULD MAKE TECHNICAL DOCUMENTATION DIRECTLY EXECUTABLE BY AGENTS
Traditional developer documentation is written primarily for humans.
A developer reads an API reference, understands the examples and converts that knowledge into software.
Agent skills insert another path:
technical documentation → structured agent knowledge → generated implementation.
This changes the role of documentation.
Hardware and software vendors can package implementation knowledge so that an AI agent consumes it directly while writing code.
For specialized platforms, this may become increasingly important because foundation models cannot be expected to contain perfectly current knowledge of every API, SDK and hardware architecture.
The quality of the skill then becomes part of the development toolchain.
··········
DOCA AGENT SKILLS FIT NVIDIA'S BROADER AGENT INFRASTRUCTURE STRATEGY
The release also connects with NVIDIA's broader push to make autonomous agents part of its infrastructure stack.
The company is developing agent systems at several layers.
DOCA Agent Skills help agents understand how to build software for NVIDIA infrastructure.
OpenShell provides a secure runtime boundary around agents.
NVIDIA Sentry and BlueField-4 provide an external hardware enforcement layer capable of monitoring agent behavior.
The components address different problems.
Skills improve what the agent knows about the infrastructure. Runtime controls restrict what the agent is allowed to do. Hardware monitoring provides an independent mechanism for detecting and stopping behavior outside defined boundaries.
Together, they point toward infrastructure designed not only to run AI models, but also to support AI agents that write, deploy and operate software.
··········
SPECIALIZED SKILLS COULD BECOME A NEW SOFTWARE DISTRIBUTION LAYER
The broader implication extends beyond NVIDIA DOCA.
Software vendors have historically distributed developer knowledge through documentation, SDKs, sample repositories and training courses.
AI coding agents create another distribution format: vendor-maintained skills specifically designed for machine consumption.
A database company could publish a skill that teaches agents its preferred schema and query patterns.
A cloud provider could package deployment, security and infrastructure guidance.
A semiconductor company could expose hardware-specific optimization rules.
An enterprise platform could provide a skill describing its APIs, permissions and workflow constraints.
Instead of waiting for those details to appear in the training data of a future foundation model, developers could give the current model an authoritative and continuously updated capability package.
··········
THE 100% RESULT HAS A NARROW BUT IMPORTANT MEANING
The headline figure requires careful interpretation.
NVIDIA is not claiming that an AI agent equipped with DOCA Agent Skills writes flawless production software with 100% accuracy.
The reported result is 100% satisfaction of checklist items across NVIDIA's 65-prompt evaluation, compared with 19% without the specialized skills.
It does not measure every possible BlueField task, long-term maintainability, production reliability or vulnerability rates.
It is also a vendor evaluation rather than an independent benchmark.
Within those boundaries, the result demonstrates a substantial difference between asking a general coding agent to reconstruct specialized platform knowledge from its model weights and giving it a dedicated, current source of hardware-aware instructions.
··········
NVIDIA IS TURNING HARDWARE KNOWLEDGE INTO AN AGENT CAPABILITY
DOCA Agent Skills represent a relatively simple architectural idea with potentially broad consequences.
A coding model does not need to contain complete and permanently current knowledge of every infrastructure platform.
It needs a reliable mechanism for acquiring that knowledge when the task requires it.
For NVIDIA, that mechanism now includes a dedicated skills layer for BlueField and DOCA development.
The company's 65-prompt evaluation shows how large the performance difference can become when an agent receives the correct specialized context: 19% checklist satisfaction without the skills versus 100% with them.
The next question is whether the same approach scales across larger projects, more complex production environments and other specialized infrastructure stacks.
If it does, agent skills could become an increasingly important interface between general-purpose AI models and the rapidly changing hardware and software systems they are expected to program.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
