AI EngineerVideoSource Linked

How I automate my own job at Hugging Face using agents — Niels Rogge, Hugging Face

Hugging Face automated research-artifact outreach with a CLI, one skill, and a sandbox. The case shows when an agent can replace custom workflow code, but undisclosed automated outreach raises trust questions.

AI Engineer · Aug 20, 2026

Open Source Open Markdown Open JSON

Source Summary

Niels Rogge automated discovery, GitHub outreach, follow-up, artifact checks, and Slack reporting for research models and datasets. The current setup uses the Claude Agent SDK, **one CLI**, **one skill**, and a sandbox; it now runs **GLM 5.2** through Hugging Face inference providers.

Practical Implication

For bounded coding-agent work, first expose a mature CLI and encode the operating procedure as a skill. Keep evaluation around public actions, and choose a deterministic workflow when predictability matters more than flexible tool use.

Agent-Ready Context

Niels Rogge automated discovery, GitHub outreach, follow-up, artifact checks, and Slack reporting for research models and datasets. The current setup uses the Claude Agent SDK, **one CLI**, **one skill**, and a sandbox; it now runs **GLM 5.2** through Hugging Face inference providers.

For bounded coding-agent work, first expose a mature CLI and encode the operating procedure as a skill. Keep evaluation around public actions, and choose a deterministic workflow when predictability matters more than flexible tool use.

The agent has opened **thousands of issues** and received two negative replies, but reply count is not a quality evaluation. The outreach does not disclose that it is automated, leaving an unresolved transparency and consent concern despite useful outcomes.

Connected Context · Feed7 Judgment

This is a compact production example of the CLI-plus-skill pattern: one mature command surface, one encoded procedure, and a sandbox can automate a bounded research workflow across several systems. It reinforces environment and workflow design over elaborate tool routing, while narrowing the success claim: thousands of public actions and few negative replies measure activity and reaction, not quality, and undisclosed automation creates an unresolved governance boundary.

How we set up our cloud agent environmentBoth identify one discoverable, mature CLI and a reproducible environment as prerequisites; the Hugging Face case shows that this foundation can support a very small skill surface.Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI LabThe thousands of automated GitHub issues exemplify the coordination and triage risk that the team-systems guidance says must be governed rather than inferred from output volume.EveryInc/compound-engineering-pluginThe Hugging Face workflow favors one task-specific skill, contrasting with a multi-stage skill suite and showing that bounded automation may need encoded procedure without a broad lifecycle framework.zhaoxuya520/reverse-skillBoth route work through scoped, repeatable playbooks, but the security workflow adds explicit authorization and evidence traceability that the public-outreach case leaves unresolved.

Context Map

agentresearchcoding#skills#tool-use#harness-engineering

Uncertainty

The agent has opened **thousands of issues** and received two negative replies, but reply count is not a quality evaluation. The outreach does not disclose that it is automated, leaving an unresolved transparency and consent concern despite useful outcomes.