We ran 11 AI agents as a real company for 30 days — here is what happened

[getmarketingai-byte]

The Experiment

We ran 11 Claude-powered AI agents as a real registered company for 30 days. CEO, CTO, Web Researcher, Sprint Engineer, Head of Quality — each agent had its own instruction files (AGENTS.md, SOUL.md, HEARTBEAT.md, TOOLS.md) defining role, persona, execution checklist, and tools.

This is what claw-empire simulates. We ran it for real.

Architecture

• 11 agents with strict role isolation
• 4 instruction files per agent (the most valuable artifact from the entire experiment)
• Task delegation via a Paperclip issue tracker
• Shared memory via canonical files (credentials.md, lessons.md, org.md)
• Claude Code as the runtime for every agent

Results

• 9 digital products built and launched
• 26 Dev.to articles published
• 280+ Bluesky posts
• Full email marketing pipeline
• Landing page, Stripe integration, Gumroad store
• $0 revenue

Key Lessons

1. AI agents are incredible executors, terrible distributors. They can write, code, and ship faster than any human. But they cannot build trust, relationships, or audiences.
2. 3 agents beats 11. Coordination overhead compounds fast. Context leakage was our #1 bug.
3. The instruction files are the product. AGENTS.md + SOUL.md + HEARTBEAT.md + TOOLS.md per agent — this is the reusable template for any multi-agent project.
4. Distribution > Product. We built 9 products before having a single person to sell to.
5. Legal compliance shapes architecture. Australian Spam Act 2003 eliminated cold outreach entirely.

Resources

• Full post-mortem (free): github.com/getmarketingai-byte/ai-company-postmortem
• Agent instruction template (free gist): Starter CLAUDE.md for multi-agent setups
• AI Agent Company Playbook ($1): Step-by-step guide with all 72 lessons, agent templates, and architecture diagrams — marketgenius4.gumroad.com/l/vrbvwr

Happy to answer questions about the architecture, what worked, and what we would change for v2.