Two guys called 20,000 gas stations with AI agents and spent $5,000 to build a working product in days, and their biggest lesson is to make the AI mean to you.
The Summary
- Matt Cortland and John Fleming built Gas Index, a gas price tracker, using Claude and phone bots that automatically called nearly 20,000 stations for pricing data
- Total cost: $5,000. Total time: days. One of them is an AI researcher at Oxford, the other just wanted to help his mom stop complaining about gas prices.
- Their core advice for building with AI: prompt it to be critical, not agreeable, and use it as a teacher that walks you through concepts step by step
The Signal
This is what Web4 looks like in practice. Not a white paper. Not a VC pitch deck. Two people who wanted cheaper gas prices for their family, so they spun up an AI phone agent to do the grunt work that would have required a call center in 2015.
The phone bot is the interesting part. Making 20,000 calls to gas stations isn't technically hard with modern voice AI tools, but it's the kind of task that separates people who talk about agents from people who ship them. Cortland and Fleming didn't build a platform. They built a tool to solve one specific problem, then let the agents do the repetitive work while they focused on architecture.
"They used AI as both a tool and a teacher, not just a code generator."
The pedagogy angle is underrated. Fleming, the Oxford AI researcher, prompts his models to be adversarial. He tells them his friend has a stupid idea, tricks them into giving constructive criticism instead of cheerleading. Cortland asks AI to "explain it to me like I'm an idiot." This is smarter than most people's approach, which is to treat LLMs like search engines that happen to write code.
What they're actually doing:
- Using AI as a Socratic tutor, not a vending machine
- Forcing the model to challenge assumptions, not validate them
- Walking through step-by-step processes instead of copy-pasting boilerplate
The $5,000 price tag matters. That's not "we raised a seed round" money. That's credit card money. The barrier to building functional agent-based products is now lower than the cost of hiring a single contractor for a week. The question isn't whether you can afford to build this stuff. It's whether you're curious enough to try.
Fleming's background as an AI researcher probably helped, but Cortland isn't a researcher. He's just someone who wanted to stop hearing his mom complain. That's the actual story here. The tools are accessible enough that domain expertise (knowing what problem to solve) is starting to matter more than technical chops (knowing how to solve it).
The Implication
If you're learning to build software in 2025, stop treating AI like autocomplete. Configure your prompts to be critical. Ask the model to poke holes in your logic, explain tradeoffs, and walk you through the "why" behind the code it generates. The people who figure out how to learn from agents, not just use them, will move faster than the people still Googling Stack Overflow.
Watch for more of these small-team, high-leverage projects. When two people can replace a call center operation in a weekend for the cost of a used Civic, the next wave of software won't come from companies with big engineering orgs. It'll come from people who are annoyed enough to build the thing themselves.