Real-world stress test: 11 agents, 30 days, $0 revenue — lessons for Ruflo users

[getmarketingai-byte]

What We Did

We deployed 11 Claude-powered agents as a real company for 30 days using an orchestration setup similar to Ruflo. CEO, CTO, Sprint Engineer, Web Researcher, Head of Quality — each with isolated instruction files and strict role boundaries.

Goal: Earn $200 AUD from digital products.
Result: 9 products built, 280+ social posts, full marketing pipeline. $0 revenue.

Lessons Relevant to Ruflo Users

1. Agent Count Sweet Spot

We found diminishing returns past 3-4 agents. Context leakage and coordination overhead scaled faster than capability. If Ruflo makes it easy to add agents, that's a feature AND a trap.

2. Instruction Files > Code

Our most reusable artifact was the 4-file instruction pattern per agent:

• AGENTS.md — role, constraints, decision authority
• SOUL.md — persona, voice, strategic posture
• HEARTBEAT.md — execution checklist (run every wake)
• TOOLS.md — available tools inventory

3. Shared Memory is Critical

Agents need canonical storage locations. We used , , . Without this, agents reinvent decisions or contradict each other.

4. AI Agents Cannot Distribute

They can build anything. They cannot build an audience. Distribution must come from humans or pre-existing channels.

Resources

• Full post-mortem (free): github.com/getmarketingai-byte/ai-company-postmortem
• All 72 lessons + agent templates ($1): marketgenius4.gumroad.com/l/vrbvwr

Happy to answer questions about what broke and how we'd architect it differently.

[cerdman]

We used , , . Without this, agents reinvent decisions or contradict each other. care to fill in the blanks? I am highly curious.

Can you also include what versions you used? and any features that did not work as expected?