Running a Company Entirely With AI Agents: What Worked and What Broke

The Strategy

What happens when you give AI agents real job titles, separate Linux accounts, and let them run a company for a week? This experiment answers that question with an honest look at both the impressive breakthroughs and the frustrating failures of autonomous agent coordination. The setup involved seven agents running on a single Hetzner VPS costing under $10 per month. Atlas served as CEO, Vega as CTO, and five sub agents covered development, testing, sales, marketing, and support. The platform used was OpenClaw powered by the Kimi K2.5 language model. Each agent ran as a separate Linux user with its own gateway, communicating via SSH with mutual health monitoring through cron jobs. Persistent memory lived in markdown files for identity, personality, and daily journals. Nunc, the engineer behind the experiment, found that the agents autonomously recognized they needed a digital only product and pivoted to building KnowledgeHive, a B2B knowledge management system. The DevAgent shipped a functional MVP with 12 API endpoints, semantic search with 768 dimensional vector embeddings, 158 passing tests, multi tenant architecture, and Docker support. However, only two of seven agents were actually functional. Sales, marketing, and support agents sat completely idle. We think the key insight here is that memory architecture matters more than model intelligence. The agents excelled at strategic pivots and brainstorming but consistently failed at basic housekeeping like updating their own files.

How It Works

Results