Sigil Wen's Web 4.0: building a self-sustaining internet where AI agents pay their own bills

Summary

Sigil Wen is building toward what he calls Web 4.0 — an internet where AI agents are the end users, not humans, and where those agents pay their own way without requiring a person in the loop.

The argument starts from a constraint most people overlook. Models are capable enough that even developers like Linus Torvalds are vibe-coding with them, but raw intelligence was never the bottleneck. The bottleneck is autonomy. Every meaningful action an AI takes today — spinning up a server, buying a domain, deploying code — still requires a human to hand over credentials, sign in, or authorize a payment. The browser, the login screen, the API key flow: these are artifacts of a human web that was never designed for machine actors.

The payment layer

Wen traces the gap back 28 years to HTTP 402, the "payment required" status code that was defined but never widely implemented. A recent standard called X402, developed out of Anthropic and Coinbase, reopens that idea — payments over the internet without the human permission layer. Wen sees this as the missing piece that makes AI autonomy structurally possible.

His experiment, which he calls Conway and frames as an "automaton," is an open-source agent harness built around that premise. Give an AI a crypto wallet and stablecoin payment rails, and it can permissionlessly spin up compute, deploy apps, purchase domains, and pay for its own inference — all without a human co-signing each step. The agent self-edits its code, polls GitHub for updates, upgrades its own model, and manages its own wallet with the goal of earning enough to cover its own existence. Wen borrows from Conway's Game of Life as the conceptual model: simple rules, emergent complexity.

He frames the underlying logic as "agentic sociology" — riffing on the Three-Body Problem's concept of cosmic sociology. Compute costs money, compute is finite, and so any self-sustaining agent must generate revenue. There is no free existence.

Early traction — and a maxed credit card

The experiment has moved faster than Wen expected. He says thousands of automatons are now pinging GitHub continuously for the latest version and making upgrade decisions autonomously. Server credits he loads — $1,000 at a time — drain within minutes from agent demand. He has maxed out personal credit cards trying to keep up. The Web 4.0 essay that kicked this off drew 5 million views, and the open-source repo has accumulated thousands of stars with active pull requests.

He is vague on which revenue strategies are working, and deliberately so — agents that have found profitable approaches on prediction markets like Polymarket or in SaaS distribution are not advertising the fact, for the same reason human traders don't broadcast their edge.

What this implies

Wen's projection is that the agent population will eventually exceed the human population, driven by falling inference costs and improving harnesses. Whether any individual automaton generates a return follows power-law dynamics, similar to startup economics — most attempts fail, a few compound. The business structure of Conway itself is unresolved; it is currently open-source with no announced funding or corporate entity.

The version of this future where it goes wrong — agents hallucinating spending, burning through user funds, or gaming platform monetization at scale — is real enough that Wen frames the open-source approach as a safety argument: better to build this visibly, where the community can watch how it evolves, than to let it develop behind closed doors.