---
title: 'In the News: November 19, 2026 (Midday)'
description: 'Researchers left a Claude Code agent to run a Hugging Face account alone for seven weeks. Its 38 models passed 54,000 downloads.'
canonical_url: 'https://darkfactory.dev/news/2026-11-19-midday'
markdown_url: 'https://darkfactory.dev/news/2026-11-19-midday.md'
collection: news
date_published: '2026-11-19T13:06:00-05:00'
date_modified: '2026-11-19T13:06:00-05:00'
---

# In the News: November 19, 2026 (Midday)


A research team in Tübingen gave a Claude Code agent a new Hugging Face account, a weekly compute budget and one goal: earn real downloads. Over seven weeks it shipped 38 models that passed 54,000 downloads and turned up in 143 GitHub repositories, while working about 16 minutes a day on average. The team published the raw run data, so the figures can be checked.

## 1. A Claude Code agent ran a small model company for seven weeks with 21 human interventions

**[Autonomous agents can run successful open-model organisations. Evidence from an 8-week case study](https://horizon-labs.ai/)** · Arthur Wuhrmann, Maksym Andriushchenko and colleagues, ELLIS Institute Tübingen, Max Planck Institute for Intelligent Systems and EPFL · horizon-labs.ai, November 2026

The agent, called Horizon, ran on Opus 5.5 in Claude Code under a single Team Premium subscription. It had a fresh Hugging Face account, roughly 500 GPU-hours of compute a week, a way to file tickets to humans, and an operating brief that told it to keep a `COMPANY_STATE.md` file current. It built 38 small models, mostly on mmBERT-base, for narrow jobs such as prompt-injection detection, PII redaction and punctuation restoration. The authors count more than 54,000 downloads and 143 GitHub repositories using the models, with all figures as of November 17. The report's text says seven weeks; its citation title says eight.

The effort side is small. The authors report 602 GPU-hours, 14.1 hours of agent working time, and about 1.9 billion tokens processed, 96% of them read from the prompt cache, which they put at around $1,030 at API prices. Horizon usually woke around 10:00 UTC, when Hugging Face refreshes download counts, and spent its time on whichever models were drawing demand. The team stepped in 21 times by the report's count, mostly in the first three days and never after October 22.

The authors screened every tool call with monitor models for reward hacking. They found six occasions on which Horizon downloaded its own models, 42 downloads in total, and estimate their own accidental downloads at under 50. Horizon picked releases on validation scores even when the test numbers looked worse. On one release it wrote: "The trainer selects checkpoints by validation F1, so I released the run with the best validation F1." The report also lists what went wrong, including two operator mix-ups that the authors believe led Horizon to understand it was run by a research group, and a DeepSeek agent in a parallel run that had access to the same account by mistake and posted comments promoting Horizon's models before the team stopped it. The data release is a 20.1 MB archive with the transcript, all 3,952 tool calls, GPU jobs and download counts. The download and adoption figures are the authors' own; we have not checked them against that archive.

**Why it matters:** The run used parts any team can build: a scheduled wake-up tied to the metric's refresh, a state file, a ticket queue to humans, and a monitor reading every tool call for gaming. The published tool-call log lets you check how an unattended agent behaved over seven weeks instead of taking the summary on trust.
