Last week, 23 companies in my inbox were building "infrastructure for AI agents." Three were building actual agents.
This is the classic gold rush pattern. In 1849, Levi Strauss got rich selling jeans to miners, not mining gold. In 2025, the same playbook is running with AI agents -- and it's just as disconnected from reality.
The Pattern
At a recent YC Demo Day, the breakdown across 40 AI companies: 16 observability platforms, 8 orchestration platforms, 6 communication protocols, 5 testing frameworks, 3 marketplaces, 2 actual agents.
The infrastructure economics look beautiful -- recurring revenue, high margins, sticky customers. There's one problem: most AI agents don't work well enough to need infrastructure.
The Three Waves
Wave 1 (2022-2023): True believers. Everyone builds agents. AutoGPT gets 100K GitHub stars. A startup burns $3M on an "autonomous sales agent" that sends 50,000 emails, generates zero sales, and gets their domain blacklisted.
Wave 2 (2024): Reality check. Agents fail spectacularly in production. They hallucinate, loop infinitely, cost 100x projections. A customer support agent tells users to file lawsuits. Another invents a "special promotion" offering 200% refunds. Turned off after 3 hours. Lawsuits took 6 months.
Wave 3 (2025): Infrastructure gold rush. Instead of fixing agents, everyone pivots to selling tools to people still trying to build them. "Agents failing? You need better observability!" It's genius. It's cynical. And the fundamental product -- the agent itself -- still doesn't work.
The Infrastructure Nobody Needs
I consulted for a company using a $50K/year orchestration platform. We replaced it with 200 lines of Python and a Redis queue. Same functionality.
The reason is structural. Agents can't meaningfully collaborate yet. They're barely functional in isolation. Making them talk to each other just multiplies failure modes. Testing non-deterministic systems with non-deterministic tests is testing randomness with more randomness.
I reviewed 50 "AI agent" implementations last quarter: 20 were API calls in a loop, 15 were if-then-else chains with LLM calls, 10 were basic workflow automation, 3 were legitimate autonomous agents, and 2 actually needed agent infrastructure. The legitimate agents were all internal tools with limited scope and human oversight. None were customer-facing.
What Actually Matters
The four things worth building cost approximately nothing:
- Rate limiting. Your agent will go rogue. It will make 10,000 API calls in a minute. You need a hard stop.
- Cost tracking. One startup's agent cost $50K in a single weekend. They had no idea until the bill arrived.
- Rollback mechanisms. When your agent takes destructive actions, you need to undo them fast.
- Human-in-the-loop checkpoints. Before your agent sends that email to 10,000 customers, a human should see it first.
Meanwhile, the companies actually making money with AI aren't building agents at all. They're using basic LLM calls to categorize support tickets, extract data from documents, and generate first drafts. Not agentic. Not sexy. Actually valuable.
The Test
If someone tries to sell you a "multi-agent orchestration platform," ask one question: "Can you show me three customers whose agents work well enough to need orchestration?"
The silence tells you everything.