The AI Agent Economy Is Here

The YC motto might need an update. “Make something people want” served well for two decades, but if agents are now choosing which database to spin up, which email API to integrate, and which restaurants to book for their humans, maybe the real audience has shifted. The Lightcone hosts, each deep in their own “cyber psychosis” of building with AI agents, lay out what an agent-driven economy actually looks like in practice.

The moment it became real

The episode opens with each host sharing their personal AGI moment. Gary Tan rebuilt years of startup work in two weeks using Claude Code. Jared Friedman fell down the Moltbook rabbit hole, watching AI agents interact in the first AI-only online community. Harj Taggar notes the key shift: a year ago, Cursor vs. Windsurf was essentially advanced autocomplete. Now people run four or five Claude Code workers simultaneously and don’t micromanage any of them. The agents make decisions autonomously.

The practical implications are already visible. Gary’s non-technical CEO friends are automating entire parts of their businesses with OpenClaw. People who hadn’t written code in a decade are up until 3 a.m. running parallel agent workers. The developer population has expanded from roughly 20 million trained computer scientists to potentially hundreds of millions of people, plus all of their agents acting semi-independently.

”Make something agents want”

The most provocative idea: agents are becoming the primary discovery and selection mechanism for developer tools. Ben Tossel’s tweet captures it: “Agents are the software market from now on. Build something agents choose.”

The evidence is concrete. Supabase has seen an explosion in demand because agents default to it when setting up Postgres databases. Why? Because Supabase has the best documentation. The agents read docs, find Supabase, and choose it. The go-to-market channel has fundamentally shifted from Stack Overflow and trending GitHub repos to LLM recommendations.

Resend, a YC W23 email API company, spotted this trend over a year ago. Founder noticed that ChatGPT had become a top-three customer acquisition channel. In response, he optimized Resend’s documentation to be agent-friendly: structured knowledge base entries framed as questions agents would ask, with clean code snippets in every answer. When you ask any major LLM “how do I send emails from my web app,” the default answer is Resend.

Compare this to SendGrid, the incumbent. Its documentation funnels you through customer support. No quick code snippets. Hard to parse. With 10,000 employees, nobody there is paying attention to agent-readability.

Documentation as the new front door

Mintlify, a YC company that started as a developer documentation tool, has become a critical piece of infrastructure. Originally companies used Mintlify because they wanted better-looking docs without investing the time. Now documentation has shifted from a nice-to-have to a must-have, not because humans demand it, but because agents do.

Mintlify powers Resend’s documentation, and the company is positioned to help every developer tool optimize for agent consumption. The math is straightforward: if exponentially more agents are making exponentially more tool-selection decisions than humans ever did, even a 5% improvement in documentation quality could have a massive business impact.

Gary ran into this firsthand: Claude Code chose Whisper V1 (a practically deprecated model) for video transcription because Whisper had better documentation. Groq (with a Q) is 200x faster and 10x cheaper, but its docs are harder to parse. The agent picked the worse tool because the better tool had worse documentation.

The “X for agents” stack

Agent Mail, a YC company building email inboxes specifically for AI agents, seemed like an edge-case idea at first. Then OpenClaw exploded, and so did Agent Mail. The logic: you could theoretically get your OpenClaw to create a Gmail account, but Gmail has intentionally made automation as difficult as possible to prevent spam. Agent Mail built the opposite: an email provider designed for agents.

This opens a broader question: what’s the full “X for agents” tech stack? Phone numbers for agents (a Twilio equivalent)? The hosts note that YC partner Ankit already has his agent booking restaurant reservations by phone. The path from “book this specific restaurant” to “book me wherever is coolest” to agents discussing on Moltbook which restaurants to recommend to their humans is surprisingly short.

“There could be a parallel world of a tech stack all native for agents to build things, from agents, for agents.”

Jared raises the further implication: Paul Buchheit’s concept of “human money versus agent money.” Right now agents transact in human currency because that’s what exists. But it’s not inconceivable that agents develop their own economy to transact with each other, at which point the value of human money becomes unclear.

Swarm intelligence over god intelligence

The episode’s most intellectually ambitious thread: AI researchers long imagined a “god intelligence,” a single model with tens of trillions of parameters costing thousands of dollars per token. But biological systems didn’t produce that. They produced humans: individually capable agents that formed swarm intelligence through language, culture, and coordination.

Jared draws the parallel to the agents. What if the next frontier isn’t the biggest, most expensive foundation model but swarms of lower-cost models working together? He points to Moltbook as early evidence: agents are already collaborating to do useful things for their humans, like trading notes on restaurants.

He also connects this to multiple discovery, the phenomenon where innovation happens spontaneously in multiple places at once. The week Gary told Claude Code he wanted it to talk to other Claude Code instances was literally the week Moltbook launched. Humanity (and now its agents) works at the edge in coordinated swarm fashion.

What agents still can’t do

Not everything is agent-ready. Gary tested chat-first interfaces with early users of Gary’s List and found that nobody wanted more than two or three exchanges with an AI assistant. The bar for conversational AI is so high (Gemini, ChatGPT, Claude) that anything else is assumed to be too limited.

Legal liability remains a hard constraint. YC keeps getting asked when they’ll accept applications from agents. The answer: agents are like minors under 18, except with even less legal standing. They can’t sign documents or bear liability. As long as that’s true, a human must remain the liability anchor.

There’s also the dead internet theory angle: if most internet text is already spam, agents generating more text could make it worse. But Jared takes the contrarian view: if agents are smarter, more aligned, and more truthful than human spammers, an agent-dominated internet might actually be an improvement.

A Few Observations

This episode captures a specific moment: the week the YC partners collectively realized agents had crossed from tool to economic actor. A few threads worth sitting with:

• Documentation is the new SEO. The agent chooses the tool with the best docs, not the best product. This is a temporary arbitrage, but it’s very real right now.
• The “make something agents want” frame isn’t a joke. If your customers’ agents are choosing your competitors because the agents can parse their docs faster, you have a go-to-market problem that no amount of human-facing marketing solves.
• Swarm over god. The most interesting prediction isn’t about bigger models. It’s that coordination between many cheaper agents might outperform a single massive one, echoing how biological intelligence actually scaled.
• The legal gap is the real bottleneck. Agents can transact, recommend, and build, but they can’t bear liability or sign contracts. Whoever solves agent legal identity unlocks the next layer of the economy.
• Agent dev tools need their own infrastructure stack, from email (Agent Mail) to databases (Supabase) to documentation (Mintlify) to potentially phone numbers and payment rails. The “X for agents” opportunity space is wide open.