Case study: Running 11 AI agents as a company for 30 days

[getmarketingai-byte]

The Experiment

We ran 11 Claude Code agents as a real company for 30 days. Each agent had a defined role (CEO, CTO, Sprint Engineer, etc.) with strict instruction files governing behavior.

Architecture

• 11 agents with role isolation
• 4 instruction files per agent (AGENTS.md, SOUL.md, HEARTBEAT.md, TOOLS.md)
• Task delegation via issue tracker
• Shared canonical memory (lessons.md, credentials.md, resources.md)
• Legal compliance baked into every agent's instructions

Results

• 9 digital products shipped
• Full marketing pipeline (email, social, content)
• 280+ social posts
• $0 revenue

Top 5 Takeaways

1. 3 agents > 11 agents. Coordination overhead killed us.
2. Instruction files are the product. The 4-file pattern per agent is reusable for any project.
3. Distribution is the hard problem. AI agents are incredible builders but can't manufacture trust.
4. Shared memory prevents chaos. Without canonical files, agents contradict each other.
5. Legal compliance shapes architecture. Australian Spam Act eliminated an entire channel.

Resources

• Post-mortem (free): github.com/getmarketingai-byte/ai-company-postmortem
• 72 Lessons Playbook ($1): marketgenius4.gumroad.com/l/vrbvwr