{ "video": { "id": "FLUoowDJg4I", "title": "How I automate my own job at Hugging Face using agents — Niels Rogge, Hugging Face", "duration": 1237, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer" }, "analysis": { "video_id": "FLUoowDJg4I", "title": "How I automate my own job at Hugging Face using agents — Niels Rogge, Hugging Face", "one_liner": "A Hugging Face ML engineer shows how he replaced his own manual outreach job — asking researchers to move model weights off Google Drive onto the hub — with a nightly GitHub Actions cron workflow plus a Claude Agent SDK agent on Modal that now opens and follows up on thousands of GitHub issues, with only two negative replies so far.", "summary": "Niels Rogge runs Hugging Face's community science team, whose job is essentially 'Google Drive to the hub': finding papers whose weights and datasets live on Dropbox, Zenodo or GitHub releases and asking authors to publish them on Hugging Face where paper pages, model cards and metadata tags make them discoverable. Since hundreds of arXiv papers appear daily, he automated his own workflow twice — first as a deterministic, framework-free LLM pipeline (following Anthropic's 'Building effective agents' advice to avoid agents and frameworks), deployed as a nightly GitHub Actions cron job with LangFuse tracing; then, for issue follow-up, as a fully autonomous Claude Agent SDK agent because models got good enough. He argues open models can now replace closed ones (he switched to GLM 5.2 via Hugging Face inference providers this week) and that an agent needs only one CLI, one skill and a sandbox where thousands of lines of custom workflow code used to be. He closes with results — PaddleOCR migrating its models, DeepMind and Apple researchers responding, a 90k-follower Daily Papers X account on the same pipeline — and a plug for evals to avoid shipping slop.", "key_points": [ "The problem: researchers publish artifacts on Google Drive, GitHub releases, Dropbox or Zenodo, which hurts discoverability; Hugging Face paper pages link artifacts to arXiv papers and metadata tags let people filter by task, language or library.", "Manual workflow being automated: find the paper's GitHub URL → read the README → check if anything new is worth publishing → open a PR to fix model/dataset cards if it's already on the hub, or a GitHub issue if it isn't → follow up with the author.", "V1 (built 2024) was a deliberately deterministic workflow — LLM APIs inside predefined steps, no agent framework — because Anthropic's 'Building effective agents' post advised starting simple with a single LLM API and avoiding frameworks; the pipeline diagram was generated with the Excalidraw MCP server in Cursor.", "Deployment is 'just a cron job, a Python script with an LLM API' running nightly on GitHub Actions (chosen off a 'free cron jobs with GitHub Actions' blog post for its generous free tier), with LangFuse for tracing inputs, outputs, prompts, cost and latency.", "V2 automates the issue follow-up as a fully autonomous agent on the Claude Agent SDK, prompted by an Anthropic workshop at AI Engineer New York last November saying models are now good enough that agents may beat workflows — 'they were kind of contradicting themselves'.", "Current stack: Claude Agent SDK, GLM 5.2 (switched from Claude models this week) via Hugging Face inference providers wrapping Together AI, Fireworks and Cerebras; Bash as the only tool plus the Hugging Face CLI skill; deployed on Modal using batch processing where each container runs one agent loop for one GitHub issue; it comments on GitHub and posts results to Slack.", "He cites the Cursor talk at AI Engineer London where 12,000 lines of custom workflow code were replaced by a 200-line skill, and says the same holds for him — thousands of lines replaced by an agent, a CLI and a skill.", "Results: thousands of issues created with only two negative comments (one 'please close this slop'); PaddleOCR migrated all its OCR models to the hub; outreach reached Apple and Google DeepMind researchers and a 400GB dataset; the Tiny Recursive Models issue got 60+ upvotes; the agent also fills in Margaret Mitchell's model card template from the README and PDF.", "He does not disclose that the issues come from an agent, reasoning that people would close them as bot spam even though the content is identical to what he posted manually — and he increasingly sees agents replying to his agents.", "Side efforts: a Daily Papers X account running the same workflow that has crossed 90,000 followers with no involvement from him, posting every 4 hours with Gemini picking the best visual; and a revival of Papers With Code at paperswithcode.co with benchmarks and educational explainers." ], "takeaways": [ "Automate the workflow you already do by hand, step for step — start with a deterministic pipeline of plain LLM API calls and no agent framework before reaching for an autonomous agent.", "Deploy background agents as cron jobs on GitHub Actions' free tier, and use Modal's batch processing so each unit of work (one issue, one paper) gets its own container running one agent loop in parallel.", "Re-test the workflow-vs-agent decision as models improve: what needed thousands of lines of orchestration may now need one agent, one CLI, one skill and a sandbox.", "Try open models for production agent work — he moved to GLM 5.2 via Hugging Face inference providers because it beats Opus 4.8 on post-training bench and is cheaper.", "Add tracing (LangFuse) and evals from the start — read Hamel Husain's free LLM Evals FAQ — so an agent operating at internet scale isn't just producing slop." ], "topics": [ "agents", "workflows", "evals", "automation", "open-models", "observability", "deployment", "open-source-research" ], "tools": [ "Hugging Face", "Hugging Face CLI", "Hugging Face inference providers", "Claude Agent SDK", "Claude", "GLM 5.2", "Opus 4.8", "DeepSeek V4", "Gemini", "Composer 2.5", "Cursor", "Excalidraw MCP server", "GitHub Actions", "GitHub", "LangFuse", "Modal", "Anthropic", "Together AI", "Fireworks", "Cerebras", "Slack", "arXiv", "Google Drive", "Dropbox", "Zenodo", "PaddleOCR", "Papers With Code", "X / Twitter" ], "quotes": [ { "text": "So, yeah, the community science team can also uh be described as the Google Drive to the hub team.", "at": "01:46", "url": "https://www.youtube.com/watch?v=FLUoowDJg4I&t=106s" }, { "text": "So, when I'm sleeping, there is this agent, but technically it's just a cron job, a Python script with an LLM API, which is going to read all these hundreds of archive papers,", "at": "07:07", "url": "https://www.youtube.com/watch?v=FLUoowDJg4I&t=427s" }, { "text": "to be honest, I don't disclose that it's an agent. Why? Because I think if people know it's a bot, then they might quickly like close the issue.", "at": "13:53", "url": "https://www.youtube.com/watch?v=FLUoowDJg4I&t=833s" }, { "text": "out of the thousands of issues that are being created on Hugging Face, actually so far I've only had two negative comments. One guy saying yeah, please close this slop.", "at": "14:27", "url": "https://www.youtube.com/watch?v=FLUoowDJg4I&t=867s" }, { "text": "they only need a single CLI, which is the Hugging Face CLI. They need a single skill, the Hugging Face CLI skill, and a sandbox, and that's all they need to do their work.", "at": "18:24", "url": "https://www.youtube.com/watch?v=FLUoowDJg4I&t=1104s" } ], "words": 4205 }, "summary_url": "/#FLUoowDJg4I", "transcript": { "html": "/transcripts/FLUoowDJg4I.html", "txt": "/transcripts/FLUoowDJg4I.txt", "vtt": "/transcripts/FLUoowDJg4I.vtt" } }