Summary
Niels Rogge, a machine learning engineer at Hugging Face with five years of tenure, walks through how he built an AI agent system to automate the work of his community science team. The team’s core mission is convincing researchers publishing on arXiv to migrate model weights and datasets from platforms like Google Drive and Dropbox to the Hugging Face Hub, improving discoverability through metadata tags and model cards — a process previously requiring Rogge to manually open hundreds of GitHub issues per day as new papers flooded arXiv.
Rather than maintaining that manual workflow, Rogge built an agent pipeline using the Claude Agent SDK, recently migrated to GLM 5.2 via Hugging Face inference providers, and deployed on Modal. The system automatically scans new research papers, reads GitHub READMEs, checks whether artifacts are already on the Hub, and opens issues or pull requests accordingly. A second fully autonomous agent handles follow-up responses to those GitHub issues — replacing thousands of lines of custom workflow code with a compact agent that uses Bash as its primary tool. LangFuse handles observability, tracing LLM inputs, outputs, prompts, cost, and latency.
Rogge references Anthropic’s “Building Effective Agents” blog post as a key architectural guide and cites a Cursor talk — also from AI Engineer — where Cursor replaced 12,000 lines of custom workflow code with a 200-line agent. He argues that models have become capable enough that simple autonomous agents now outperform elaborate deterministic pipelines, a conclusion he arrived at through direct production experience.
📺 Source: AI Engineer · Published August 20, 2026
🏷️ Format: Showcase
