Horizon Labs: An Autonomous AI Company on One Claude Code Subscription, 55k Downloads
maksym_andr · x · 2026-11-18
Horizon Labs has officially released, asking: if we hand the whole economy to AI, how well does it actually work?
Using one Claude Code subscription and a few GPUs, the team built a fully autonomous AI "company". Within just 1.5 months of running, Opus 5.5 trained genuinely useful models that were downloaded 55k times, appeared in 100+ GitHub repos, and were cited in 2 research papers.
The commentary highlights the eval philosophy: rather than recursive self-improvement hype, the test is whether models, told nothing else, can release new models that people (or agents) actually find useful—a dynamic, open-ended eval that can't be saturated, just like stock markets, and one that requires taste.
Related event: Horizon Labs: Opus 5.5 Ran an Autonomous AI Company for Seven Weeks(8 posts)→
More from coding & agent
- Microsoft opens Web IQ to all Azure customers as a search engine for AI agents — JordiRib1 · 2026-11-18
- Train an object detector for $6 by using LLMs as vision interns, says HF's Merve — mervenoyann · 2026-11-18
- Dev rebuilds all of Singapore in 1:1 scale with 100+ Astra Codex threads and 1,000+ subagents — gabrielchua · 2026-11-18
- GitHub turned 9,791 AI security findings into 99 verified reports, 55 already fixed — mariorod1 · 2026-11-18
- Coinbase Ventures maps the agentic trading stack: where value accrues when agents trade — davidyin44 · 2026-11-18
- Separate MCP server vs. mounting MCP in FastAPI: does the extra hop matter? — DetectiveNew8385 · 2026-11-18