⚡

Developer 3s Key Decision Metrics

TL;DR Verdict
🏆 Benchmark Rank72s 必看
Vendor 65-prompt eval: 19% → 100% ch
Full Leaderboard→
💰 Cost & Plans比价
API & Subscription
Compare Plans↗
🛠️ Ecosystem & IDERules
claude-code · openai-codex · doca-agent-skills
Setup IDE Rules↗

On 2026-10-01 NVIDIA published DOCA AI agent skills on NVIDIA/skills: SKILL.md packs with real API signatures, hardware capability checks, and build constraints for Flow, GPUNetIO, PCC, RDMA, and more. Vendor 65-prompt eval: 19% checklist items without skills vs 100% with; RDMA demo used 73% less handwritten code and 46% fewer hardware commands. Installable into Claude Code, Codex, and other coding agents.

Key Takeaways

  • ✓Blog 2026-10-01 + https://github.com/NVIDIA/skills
  • ✓SKILL.md: real APIs, hardware capability manifests, build/preflight contracts
  • ✓Vendor 65-prompt eval: 19% → 100% checklist; top failures without skills documented in blog
  • ✓RDMA demo: −73% handwritten code, −46% hardware commands (vendor)
  • ✓Install into Claude Code/Codex-class agents; keep human gates on firmware writes
🧭

Finished reading? Explore benchmark rankings & pricing

Real-world SWE-bench scores & $20/mo vs API cost break-even calculator

🏆 Leaderboard→💰 Compare Plans↗
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
📝
Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills
developer.nvidia.com/blog/build-applications-on-nvidia-bluefield-faster-with-nvidia-doca-agent-skills/
↗
🐙
NVIDIA/skills — Agent Skills for NVIDIA products
github.com/NVIDIA/skills
↗
📖
NVIDIA DOCA SDK documentation
networking-docs.nvidia.com/doca/sdk/
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points

General coding agents invent plausible DOCA APIs/flags, skip hardware capability checks, and fail at link time—or worse, touch firmware without preflight. NVIDIA’s 2026-10-01 post frames the gap: without a machine-readable contract, agents guess from general training data.

Architecture Highlights & Internals

DOCA AI agent skills ship as SKILL.md packs with real signatures, capability manifests, build constraints, and failure mitigations across Flow, GPUNetIO, PCC, RDMA, and more. Firmware-class changes require maintenance windows, OOB assumptions, rollback, and cold power-cycle notes. Repo: NVIDIA/skills; install into Claude Code, Codex, and other Agent Skills hosts.

Authoritative Benchmarks & Measured Scores

Vendor-only numbers from the blog—do not invent external benches. 65 real prompts graded on required-answer checklists: 19% → 100% with skills; without-skills failure counts include API/flag misuse 59/65 and missing hardware checks 46/65. RDMA demo on BlueField-3: 189 vs 695 handwritten lines (−73%), 20 vs 37 hardware commands (−46%). Results depend on vendor prompts/graders.

Developer Hands-on Guide

Pull the relevant skills from NVIDIA/skills, install into your coding agent, verify device support before codegen, use pkg-config for link flags, and keep human change control for production DPU firmware writes.