Your Coding Agent Should Do AI System Engineering — Ben Burtenshaw, Hugging Face

2026-05-21 · Source: AI Engineer · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, long

Summary

Ben Burtenshaw from Hugging Face proposes that coding agents can tackle complex AI systems engineering problems, moving beyond basic coding tasks. The talk outlines three progressively autonomous applications: interactively writing optimized CUDA kernels, zero-shot fine-tuning of large language models (LLMs) on Hugging Face, and establishing multi-agent auto research labs. For CUDA kernels, agents can generate valid and optimized code, with Hugging Face's "kernels" library facilitating distribution and compatibility. An example showed a 94% speedup for a Qwen 3 8B kernel on an H100 GPU. Agents can also fine-tune models like Qwen 3 6B on specific datasets. The most advanced application, AutoLab, involves a distributed multi-agent system (researcher, planner, workers, reporter) for automated AI research, integrating with GitHub, HF jobs, and the "Trackio" dashboard for metrics.

Key takeaway

For AI Engineers seeking to optimize model performance and automate research, you should integrate coding agents into your workflow. Agents can generate custom CUDA kernels for significant speedups, like the 94% observed for Qwen 3 8B on H100, and fine-tune LLMs efficiently. Explore Hugging Face's "kernels" library and "AutoLab" multi-agent framework to scale your engineering efforts and tackle complex system-level challenges, leveraging open primitives for greater control and adaptability.

Key insights

Coding agents can perform complex AI systems engineering tasks, from kernel optimization to automated research.

Principles

• Agents can write valid, optimized CUDA kernels.
• Memory, not compute, is often the GPU bottleneck.
• Open primitives and tools enhance agent control.

Method

The AutoLab multi-agent research setup involves a researcher finding hypotheses, a planner queuing jobs, workers implementing training scripts, and a reporter monitoring progress via Trackio.

In practice

• Use Hugging Face "kernels" for custom kernel distribution.
• Employ "upskill" to evaluate agent performance with open models.
• Try AutoLab for automated AI research experiments.

Topics

• Coding Agents
• AI Systems Engineering
• CUDA Kernels
• LLM Fine-tuning
• Multi-Agent Research
• GPU Optimization

Best for: AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation — Agentic reinforcement learning can significantly enhance LLMs' intrinsic CUDA kernel optimization capabilities.
• OpenAI Agents SDK Crash Course (with Hugging Face Models) — OpenAI Agents SDK simplifies AI agent creation with minimalist primitives, supporting both OpenAI and Hugging Face models.
• An AI agent coding skeptic tries AI agent coding, in excessive detail — Modern AI agents, guided by explicit `AGENTS.md` files, can generate highly optimized, performant code across multiple languages.
• Open Source Coding Agents: The Future of AI-Driven Software Development — Open source coding agents enable autonomous, scalable software development by orchestrating multi-agent workflows and offering model flexibility.
• TensorBench: Benchmarking Coding Agents on a Compiler-Based Tensor Framework — Repository-level coding benchmarks require difficult tasks with robust, behavior-preserving evaluation to separate capability from noise.
• 🥇Top AI Papers of the Week — AI advancements focus on novel computing, efficient memory, robust agents, and specialized models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.