Hugging Face refactors the 🤗 Kernels project, upgrading GPU kernels from scattered scripts to first-class Hub citizens: a new kernel repository type, all kernels listed under huggingface.co/kernels, with a directly visible accelerator (CUDA/ROCm/XPU), OS, and backend-version support matrix, source Git SHA1 embedded in compiled artifacts, paired with Nix hermetic build, any release version can be independently reproduced, bidding farewell to the "I built it different from you" mystic problem. Security is raised to the protocol layer: only trusted publishers are allowed to release by default, with Sigstore cosign kernel signing + GitHub Actions workflow dual verification — even if an account is compromised, the attacker cannot sign out a malicious version. This ports OCI supply-chain security experience to the ML kernel ecosystem. Paving the way for Agent engineering: the kernels and kernel-builder CLIs are cleanly split, the latter intentionally "agent-optimized" (non-interactive, parseable output), paired with backend skills files, Agents can end-to-end scaffold, build, benchmark, and iteratively optimize a flash-attention-class kernel; a new Torch Stable ABI gives about 2 years of backward compatibility, Apache TVM FFI lets the same kernel run across PyTorch / JAX / CuPy. flash-attention, Fused MoE, MLA decoding and other high-frequency kernels are now install-and-use, trustworthy, and reproducible — no more "upload to PyPI + write CMake + write gotcha docs". The closed loop of Agent auto-generating and verifying kernels has a real landing foundation; the "infrastructure standardization" of the LLM training-inference stack in the second half of 2026 has just begun running.