Your Coding Agent Should Do AI System Engineering
summary
Claims include an RMSNorm kernel delivering 1.88x speedup on H100 and a finetuned Qwen3 0.6B reaching 35% on LiveCodeBench; agents use Claude Code and file-based skills to write, benchmark and publish CUDA kernels to the Hugging Face Hub, run HF jobs for end-to-end fine-tuning, and orchestrate a multi-agent AutoLab that pushes experiment metrics to a Trackio dashboard while branches and patches capture reproducible changes.
AI-written summary · about
chapters (11)
- 0:00Introduction to AI Systems Engineering
- 1:59Boss 1: Writing and Distributing CUDA Kernels
- 3:48Efficiency in Deep Learning
- 6:08Using Skills for Agentic Workflows
- 8:37Benchmarking and Evaluating Skills with Upskill
- 9:26Boss 2: End-to-End Fine-tuning of LLMs
- 10:16Boss 3: Multi-Agent Auto Research Labs
- 12:09Architecture of the Multi-Agent Research System
- 13:40Implementing the Research Agent in OpenCode
- 15:28Monitoring Experiments with Trackio
- 16:45Final Takeaways and Conclusion
related talks
30mAgents Building Agents
27mSoftware engineering with GenAI
26mAI Agents as Your Co-Developer
30mBeyond the Hype: Using AI Coding Agents Today
01h 53mMastering Agentic Coding and GPUs
16mCoding Reinvented How AI Agents Are Shaping the Future of Software Development
30mHow AI Agents Will Change the Future of Programming
26mFrom autocomplete to agents: AI coding assistance state of play