A couple just called 20,000 gas stations with an AI agent they built in a weekend, and the real story isn't the gas prices.

The Summary

  • Matt Cortland and John Fleming built Gas Index using Claude Code, an AI phone agent that called 20,000 US gas stations to collect real-time pricing data
  • The bot, named Bobby, interacts naturally with station attendants, revealing how people react to AI agents in real-world interactions (most answer, some hesitate, a few swear)
  • This is vibe-coding in production: two people with no product infrastructure shipped a working data collection system in days, not months

The Signal

This is what the agent economy looks like when it escapes the demo. Cortland and Fleming didn't build an app that scrapes GasBuddy or calls an API. They built an agent that talks to humans, thousands of times, collecting data that doesn't exist in structured form anywhere else. That's the shift. The infrastructure layer of Web4 isn't always APIs and smart contracts. Sometimes it's a bot with a Texas drawl making 20,000 phone calls because that's still how gas stations share their prices.

The speed matters. Fleming is a postdoc at Oxford, Cortland is a consultant. Neither are professional developers. They used Claude Code to ship this over a weekend. That's the vibe-coding thesis playing out: AI tools are compressing the distance between idea and execution. What used to require a team, funding, and months of coordination now happens in 72 hours because the LLM writes the code and the voice agent handles the tedious work.

The human reactions are the tell. Most attendants just answered. Some paused. A few cursed. But the system worked. Bobby got data from 20,000 locations because the interaction was smooth enough that people didn't hang up. That's the real test. Not "can an agent make a phone call" but "will a busy gas station attendant in Kansas City treat this bot like a legitimate request?" Apparently yes.

The Implication

If two people can deploy 20,000 agents to collect pricing data in a weekend, what stops someone from doing this at scale for every fragmented, unstructured data source in the economy? Real estate listings. Local restaurant menus. Small business inventory. The next wave of companies won't be building better dashboards. They'll be building agents that go get the data that doesn't exist yet. Watch for the backlash when businesses realize bots are calling them en masse. And watch for the infrastructure plays that make this kind of deployment trivial.

Source: Business Insider Tech