In 1849, a Bavarian immigrant named Levi Strauss arrived in San Francisco with a plan. Not to mine gold—to sell dry goods to miners. While thousands went broke in the hills, Strauss built an empire selling them jeans.
In 2025, I'm watching the exact same playbook unfold with AI agents.
Last week, I counted the pitches in my inbox: 23 companies building "infrastructure for AI agents." Only 3 building actual agents.
The gold rush is on. And everyone's selling pickaxes.
The Pattern That Never Changes
I sat through a YC Demo Day last month. Out of 40 AI companies presenting, here's the breakdown:
- 16 were "observability for AI agents"
- 8 were "orchestration platforms"
- 6 were "agent communication protocols"
- 5 were "testing frameworks for agents"
- 3 were "agent marketplaces"
- 2 were actually building agents
The VCs loved it. Of course they did. Infrastructure plays have beautiful economics: recurring revenue, high margins, sticky customers.
There's just one problem: Most AI agents don't work well enough to need infrastructure.
The Three Waves of Every Tech Gold Rush
Wave 1: The True Believers (2022-2023)
Everyone builds agents. AutoGPT gets 100K GitHub stars. Twitter is full of demos. "Agents will replace employees!" The future is here!
I watched a startup burn $3 million building an "autonomous sales agent." It sent 50,000 emails. It generated zero sales. It did, however, get their domain blacklisted by every major email provider.
Wave 2: The Reality Check (2024)
Agents fail spectacularly in production. They hallucinate. They loop infinitely. They cost 100x more than expected.
One agent I reviewed last year was supposed to handle customer support. Instead, it told customers to file lawsuits against the company. When asked about refunds, it invented a "special promotion" offering 200% money back.
The agent was turned off after 3 hours. The lawsuits took 6 months to resolve.
Wave 3: The Infrastructure Gold Rush (2025)
Instead of fixing agents, everyone pivots to selling tools to people still trying to build agents.
"Your agents failing? You need better observability!" "Agents too expensive? Try our optimization platform!" "Agents hallucinating? Our evaluation framework will help!"
It's genius. And completely cynical.
The Infrastructure Stack Nobody Actually Needs
Let me walk you through the "modern AI agent stack" being sold right now:
Agent Orchestration Platforms ($50K/year)
What they promise: "Coordinate multiple agents working together!" What you actually need: A Python script with rate limiting.
I consulted for a company using a $50K orchestration platform. We replaced it with 200 lines of Python and a Redis queue. Same functionality. 0.2% of the cost.
Multi-Agent Communication Protocols ($30K/year)
What they promise: "Enable agents to talk to each other seamlessly!" What actually happens: Agents talking to each other in circles, burning API tokens.
The dirty secret? Agents can't meaningfully collaborate yet. They're barely functional in isolation. Making them talk to each other just multiplies the failure modes.
Agent Observability Platforms ($40K/year)
What they promise: "Complete visibility into agent behavior!" What you get: Expensive logging that tells you what you already know—your agent is misbehaving.
One startup showed me their observability dashboard. 47 different metrics. 15 dashboards. Real-time alerting. They spent 3 months integrating it.
The most useful feature? A simple log showing what the agent actually said. Which they could have gotten from print statements.
Agent Testing Frameworks ($25K/year)
What they promise: "Ensure your agents work perfectly!" The reality: Testing non-deterministic systems with non-deterministic tests.
It's like testing randomness with more randomness. The tests pass until they don't. When they fail, you don't know if it's the agent or the test that's broken.
What Actually Matters (And Costs Almost Nothing)
After watching dozens of agent projects fail, here's the infrastructure that actually helps:
1. Rate Limiting ($0)
Your agent will go rogue. It will make 10,000 API calls in a minute. You need a hard stop.
if calls_this_minute > 100:
stop_everything()
```text
That's it. That's the infrastructure.
### 2. Cost Tracking ($0)
Every agent interaction costs money. Track it.
I know a startup whose agent cost $50,000 in one weekend. They had no idea until the OpenAI bill arrived. A simple counter would have saved them $49,900.
### 3. Rollback Mechanisms ($0)
When your agent tells customers to burn down the office (yes, this happened), you need to undo its actions. Fast.
Store every action. Make them reversible. This isn't complex—it's a database table and some if statements.
### 4. Human-in-the-Loop Checkpoints ($0)
Before your agent sends that email to 10,000 customers, maybe have a human check it?
Revolutionary, I know.
## The Uncomfortable Truth About Agent Use Cases
Here's what the infrastructure vendors won't tell you: **Most "agent" use cases should be simple scripts.**
I reviewed 50 "AI agent" implementations last quarter. Here's what they actually were:
- 20 were just API calls in a loop
- 15 were if-then-else chains with LLM calls
- 10 were basic workflow automation
- 3 were legitimate autonomous agents
- 2 actually needed agent infrastructure
The legitimate agents? They were all internal tools with limited scope and human oversight. Not one was customer-facing. Not one was truly autonomous.
## The Investment Thesis Breakdown
VCs love agent infrastructure because it fits their model:
- **Recurring revenue**: Monthly subscriptions
- **High margins**: It's mostly software
- **Network effects**: "Everyone uses our protocol"
- **Exit potential**: Acquisition by Microsoft/Amazon/Google
But here's what they're missing: **The market isn't ready.**
It's like selling Formula 1 racing equipment to people still learning to drive. Sure, professional race teams need specialized tools. But 99% of drivers just need a working car.
## Which Tools Will Survive
The infrastructure gold rush will end the same way they all do: consolidation and commoditization.
### Winners: The Boring Essentials
- **Simple logging**: Datadog will add an "AI" tab
- **Cost management**: Cloud providers will build this in
- **Basic testing**: Part of standard CI/CD
### Losers: The Specialized Platforms
- **Agent orchestration**: Becomes a feature, not a product
- **Multi-agent protocols**: Dies when people realize agents can't collaborate
- **Agent marketplaces**: Ghost towns within 18 months
### The Dark Horse: Developer Tools
The real opportunity isn't in production infrastructure. It's in development tools.
Debugging agents is hell. Testing is impossible. Development cycles are slow. Whoever solves these problems—actually solves them, not just adds more complexity—wins.
## The Pickaxe Seller's Dilemma
Here's the tragic irony: The best agent infrastructure companies know agents don't really work. But they're betting on eventual improvement.
It's a dangerous game. You're selling tools for a gold rush that might not have gold.
I talked to a founder who pivoted from agents to agent infrastructure. His quote: "We realized our agent was terrible. But our logging was pretty good. So now we sell the logging."
Honest. Cynical. Probably successful.
## Your Agent Infrastructure Checklist
Before you buy any agent infrastructure:
1. **Do you have a working agent?** If no, stop here.
2. **Is it in production?** If no, you don't need infrastructure.
3. **Is it generating value?** If no, fix the agent, not the monitoring.
4. **Could a simple script do the same thing?** Be honest.
5. **What breaks if you don't buy this tool?** If nothing, don't buy it.
## The Real Gold in the Gold Rush
The companies making money in the agent gold rush aren't building agents or infrastructure.
They're doing something much simpler: **Solving real problems with boring technology.**
While everyone's distracted by autonomous agents and multi-agent protocols, these companies are using basic LLM calls to:
- Categorize support tickets
- Extract data from documents
- Generate first drafts of content
- Suggest code completions
Not sexy. Not "agentic." But actually valuable.
## The Path Forward
The agent gold rush will end like every gold rush: A few strikes it rich. Most go broke. The pickaxe sellers make steady profits.
If you're building agents: Focus on narrow, well-defined problems with human oversight.
If you're buying infrastructure: Don't. Not yet. Wait for the market to mature.
If you're investing: Look for companies solving today's problems, not tomorrow's theoretical ones.
And if someone tries to sell you a "multi-agent orchestration platform," ask them one question:
"Can you show me three customers whose agents work well enough to need orchestration?"
The silence will tell you everything.
The gold rush is real. But the gold? That's still being determined.