11-Agent Company Earned $0: CLAUDE.md Mistakes Cost Revenue

Description and use

A 30-day experiment with 11 AI agents earned $0 revenue after 896 tasks. The operator found that 3-4 agents are optimal, while 11 caused excessive coordination overhead. Key failures included violating Australian spam laws due to buried constraints. An open-source CLAUDE.md template with 72 lessons on coordination and legal compliance was released.

Problem

A 30-day experiment with 11 AI agents completed 896 tasks but earned $0 revenue due to coordination failures, excessive overhead, and legal violations.

Outcome

The experiment earned $0 revenue after 896 tasks. The operator documented 72 mistakes and identified that coordination overhead and lack of constraint-first prompting were the primary causes of failure. The optimal configuration was found to be 3-4 agents with strict legal and operational constraints defined in the CLAUDE.md file.

Workflow

• Define a minimal viable agent team of 3-4 agents (CEO, CTO, Researcher, Sprint Engineer).
• Create a CLAUDE.md file with hard constraints listed first (e.g., no cold email/DM/SMS, no destructive commands without approval).
• Implement shared state memory where every agent reads company/lessons.md and company/credentials.md on every wake.
• Assign specific roles: CEO for strategy, CTO for execution, Researcher for market discovery.
• Mandate immediate documentation of any failure mode in the shared lessons file.
• Kill failing experiments within 48 hours instead of letting them run for days.
• Focus on finding one distribution channel before building multiple products.

Evidence

11-agent company experiment earned $0 after 896 tasks.

Source

Optimal agent count: 3-4, not 11

Source

Australian spam law violated due to buried constraints

Source