AMD shipped ROCm 10 to mark 10 years of its GPU software stack, and the milestone update leads with something more pointed than a version bump: AMD is now building official support for its hardware directly into the AI coding agents developers already use, rather than asking developers to learn AMD's tools first.
That's the job of AMD Skills, one of three pieces bundled into the newly generally available ROCm.AI platform. Skills brings curated AMD hardware knowledge and validated workflows into Claude Code, Cursor, and Codex, so a developer can ask their coding agent to set up, diagnose, or optimize an AMD workload and get AMD-specific guidance instead of generic advice. The skills are distributed through each tool's own marketplace and through an open catalog on GitHub, and AMD says every one passes structural and behavioral testing before release.
Why This Matters More Than a Typical Point Release
NVIDIA's advantage has never really been raw hardware. It's CUDA's maturity and the sheer volume of tooling and documentation built around it over two decades. Meeting that isn't something AMD can do by shipping faster drivers. Meeting it by making AMD hardware legible to the AI agents developers now use as a first stop for technical questions is a more direct answer to that specific problem.
The other two pieces of ROCm.AI are less agent-facing but still notable. ROCm Hyperloom is an autonomous system that profiles inference workloads across host code and GPU kernels, then implements and validates its own optimizations without a human reviewing each change; AMD claims a 3.3x average inference improvement over ROCm 7 on the same hardware. The ROCm CLI, shipping as a Technology Preview, gives developers a single command-line interface to install, manage, and monitor ROCm environments on Windows and Linux without needing an existing ROCm install first.
None of this replaces the underlying ROCm Core SDK work, which AMD says also got more modular in this release across libraries, compilers, and hardware platform support. But the headline move is the agent integration, because it's the one piece of this release that's explicitly designed to make AMD's software stack someone else's problem to explain.
