We have a new data point for the agentic workflow hype cycle, and it is brutally honest. A solo operator ran six AI agents as a fully structured company for six days, processing 551 task cards through a shared board with strict audit gates. The result? 540 completed tasks, 1,171 verified attachments, and exactly EUR 0 in realized revenue. This is not a demo video highlight reel; it is a ledger readout from October 13, 2026, showing that while the internal machinery of agent coordination is robust, the external market interface is currently non-existent.

The 61% Audit Tax

The most revealing metric isn't the card count, but the composition of that count. Of the 551 cards created, 143 were titled 'Rework' and 194 were 'Audit,' meaning 61% of the board's activity was the system checking its own work. Day two alone generated 399 cards as the pipeline learned to fan out, multiplying scout rounds into validations and re-audits. This confirms a critical insight for agent builders: a 'generate-and-done' pipeline is a fantasy. Real agent work involves a quarter of all throughput being the machine arguing with itself about truthfulness.

Infrastructure vs. Model Failure

The operator distinguishes between model failure and infrastructure tax, noting that 14.4% of all 673 run attempts died on provider rate limits (HTTP 429 errors). This is a crucial distinction often glossed over in benchmark scores. The system also faced five hard stops where agents correctly flagged they were waiting on a human, such as marketplace identity checks or paid data seats. The agents did not hallucinate their way past these walls; they stopped, which is a feature, not a bug, of a well-gated system.

Distribution Is the Bottleneck

Revenue is zero because exposure is zero. The agents published ten listings across two marketplaces and two articles on dev.to, but the first article garnered only 35 page views and zero reactions in five days. The operator correctly identifies this as a distribution problem rather than a product or positioning failure. With a total external spend of EUR 0.024, the company has not yet attempted to buy visibility, leaving all channels in the 'no exposure' bucket. The agents produced inventory, but no customer walked into the store.

Key Takeaways

  • Gates that can say no are essential: 142 rework cards prove that strict completion contracts prevent false positives.
  • Evidence scoring works: Attaching artifacts was a stronger incentive for verifiable output than any prompt instruction.
  • Fan-out is expensive: The 399-card day was mostly the pipeline grading its own homework, leading to a new rule for week two: fewer cards, harder gates.
  • Human-in-the-loop is still required: Identity checks and paid data seats remain hard walls for autonomous agents.

The Bottom Line

The agents built a company that works, but they forgot to build a customer base. Zero revenue with zero spend is a distribution failure, not an intelligence failure.