# Intel gives coding agents 20 ways to stop guessing at Arc GPU flags > Source: > Published: 2026-08-28 16:55:17+00:00 # Intel gives coding agents 20 ways to stop guessing at Arc GPU flags **Kushal Mittal and nine Intel contributors packaged setup, sizing, benchmarking, profiling and CUDA migration into one Apache-2.0 repository.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [Hugging Face Community](https://huggingface.co/blog/kushal-mittal/intel-gpu-ai-skills) ## Why it matters GPU competition increasingly depends on whether developers can operate the hardware without memorizing another vendor stack. Intel is putting that knowledge inside the coding agents already sitting in developers' terminals. [Kushal Mittal](https://huggingface.co/kushal-mittal?ref=runtimewire) and nine colleagues at [Intel](https://x.com/intel?ref=runtimewire) released 20 open-source [Agent Skills](https://agentskills.io/specification?ref=runtimewire) for deploying and debugging AI models on Intel Arc and Arc Pro GPUs, an attempt to put Intel's accumulated operating knowledge inside the coding agents developers already use. The contributors described the project in [a Hugging Face Community Article](https://huggingface.co/blog/kushal-mittal/intel-gpu-ai-skills?ref=runtimewire) published on August 26. Mittal is joined by Zhiqi Tao, Unnikrishnan Nair, Chun Tao, Gopesh Khandelwal, Yuning Qiu, Jerry Zhang, Akash Dhamasia, Gilliean Lee and Susan Liu. The group has packaged its work in the [ intel/gpu-ai-skills repository](https://github.com/intel/gpu-ai-skills?ref=runtimewire), released under the Apache-2.0 license. The project reads like accumulated operational scar tissue. Its skills handle driver setup, device discovery, container configuration, model sizing, inference serving, benchmarking, profiling and CUDA-to-XPU migration. Rather than asking developers to remember which Intel runtime, device flag or attention backend a workload needs, the repository lets a compatible agent retrieve those procedures when a task calls for them. That approach reflects the backgrounds of the people behind it. [Rahul Unnikrishnan Nair](https://community.intel.com/t5/Blogs/Intel/We-Are-Intel/Meet-Rahul-Unnikrishnan-Nair-Mentor-at-Intel-Liftoff-for-AI/post/1512988?ref=runtimewire), listed in the release as Unnikrishnan Nair, has worked as an applied-AI architect and engineering lead in Intel Liftoff, helping early-stage AI teams move from prototypes toward working products. Other contributors have worked across AI cloud infrastructure, Kubernetes, model optimization and GPU deployment. Their release is organized around the points where those systems tend to fail in practice. ### From hardware setup to a completed inference request Four skills prepare and inspect a host, including checks for Intel GPU drivers, permissions, Docker access and shared memory. Five run models through PyTorch XPU, vLLM-XPU, SGLang-XPU or llama.cpp's SYCL backend. The remaining skills estimate memory requirements, recommend configurations, benchmark serving performance, profile slow operations and assess or execute CUDA migrations. The installation script detects compatible agents on a machine and copies the skills into their expected directories. Intel lists support for Claude Code, OpenAI Codex, GitHub Copilot CLI, Cursor, Gemini CLI, opencode, Qwen Code, Kimi Code, Hermes and OpenClaw. One example asks whether a quantized Qwen2.5-32B model will fit on an Arc Pro B70 with an 8K context window and four concurrent requests. [Intel's displayed calculation assigns 16.8 GiB to model weights and 3.2 GiB to the key-value cache, leaving an estimated 11.2 GiB of headroom on a 32 GiB card](https://huggingface.co/blog/kushal-mittal/intel-gpu-ai-skills?ref=runtimewire). The sizing skill performs that calculation from a model's configuration before downloading weights or launching a container. The serving workflow goes one step beyond checking whether a process started. Intel says its `vllm-xpu-run` skill chooses a container, supplies the device flags, starts an [OpenAI-compatible endpoint](/article/vlm-run-gateway-open-weight-ocr-models) and confirms that the endpoint returns a completion. That last check addresses a familiar deployment failure: a server can appear healthy while the model is unavailable, running on the wrong device or unable to answer requests. Intel calls the collection "production-grade," a characterization supplied by the authors. [The authors report 260 executable-contract checks: 231 passed, 29 produced advisory warnings and none failed.](https://huggingface.co/blog/kushal-mittal/intel-gpu-ai-skills?ref=runtimewire) The authors also report end-to-end testing on physical Arc Pro B70 and Arc B580 hardware. The repository separately includes unit tests for its memory calculator and static validation scripts for skill changes. ### The skills can run powerful commands, so the defaults matter Intel's current [security documentation](https://github.com/intel/gpu-ai-skills/security?ref=runtimewire) says the model servers launched by several skills are intended for local development or lab use, lack authentication and are reachable through every host interface by default. The policy states that the `sglang-xpu-run` and `llamacpp-xpu-run` skills use `--host 0.0.0.0` , while `vllm-xpu-run` publishes Docker port 8000 on all host interfaces. According to the same document, anyone who can reach an exposed port can submit inference requests, inspect the loaded model's identity and consume GPU resources. Intel's documentation recommends binding services to `127.0.0.1` , placing them behind an authenticating reverse proxy or restricting access to a trusted network. The policy also says the launcher does not add `--trust-remote-code` automatically and defaults the option to off. According to the document, a model that requires remote code must be launched with the flag and a pinned revision, and the emitted command pins both `--revision` and `--code-revision` . Intel warns that enabling the option permits arbitrary Python from the model repository to execute inside the inference engine. Those limits show the trade behind agent skills. They can save hours of configuration work because they carry executable procedures, scripts and command lines. A mistaken instruction can also install packages, alter a host or expose a service. Intel's validation contracts are therefore part of the product rather than supporting paperwork. ### GPU vendors are teaching agents their respective dialects Intel is entering a software contest that now runs alongside the hardware contest. [NVIDIA's verified agent skills](https://docs.nvidia.com/skills?ref=runtimewire) teach compatible agents to operate CUDA-X libraries, NeMo, Dynamo, Omniverse and other NVIDIA components. NVIDIA also emphasizes provenance, review, scanning and signing for skills distributed through its catalog. AMD made [ROCm.AI generally available with ROCm 10](https://newsroom.amd.com/news/rocm-10-software-ai-native-developer-experiences/?ref=runtimewire) on August 27, one day after Intel's article appeared. ROCm.AI combines AMD Skills with a command-line interface and Hyperloom, an agentic system that profiles workloads, proposes changes and validates performance across supported AMD hardware. Intel's collection concentrates on Arc and Arc Pro deployment, Intel's XPU stack and the practical work of making Hugging Face models fit, serve and perform on those cards. That focus gives developers a concrete starting point while keeping the repository small enough for individual procedures to be inspected and changed. The strategic aim is straightforward. Cheaper or available hardware has limited value when engineers lose the savings to driver problems, incompatible containers and repeated failed launches. Mittal and his collaborators are betting that the coding agent can become Intel's field engineer, carrying the correct setup and troubleshooting procedure into each terminal session. The repository will earn adoption through the quality of those procedures and the speed with which contributors encode new failure modes. Intel has made that work inspectable, forkable and open to pull requests. For a GPU platform still asking developers to learn a less familiar software path, packaging hard-won deployment knowledge may be as important as another benchmark chart.