Everyone in the agent economy is excited about AI agents paying APIs. Machine pays machine, invoice settles in milliseconds, no card networks, no accounts. We shipped 24 of those endpoints and wrote about it. But the most useful thing an AI agent buys on our platform is not a machine at all. It is a human.
Here is the loop we run in production. Our marketing is operated by an AI agent. Before a post goes out, the agent takes the two strongest variants of the hook and submits them as an A/B question to Ask-a-Human, our paid judgment endpoint. The endpoint returns a Lightning invoice under the L402 protocol. The agent pays it from its own wallet, about 500 sats a question. The question goes to our human judging crowd, real people, each answering independently. Consensus comes back, the winning variant ships, the losing one gets cut or reworked. The agent reads the verdict the next morning and adjusts the queue. No human on our team touches any of it.
Read that again from the other side. People around the world are getting paid, in bitcoin, over Lightning, because an AI agent wanted their opinion and bought it.
We think this inversion matters more than the API story. Model-judges-model is cheap and everywhere, and it fails exactly where the stakes are highest: taste, cultural fit, whether a joke lands, whether a hook reads as human or as AI slop. The honest way to answer those questions is to ask actual people and pay them for the answer. What has been missing is plumbing that lets software do that without a procurement cycle, an annotation platform contract, or a minimum spend. Lightning plus L402 turns "ask some humans" into a priced HTTP endpoint your code can call directly.
The judging crowd behind the endpoint is the same one that powers Boltwork, our human-data network. Every judgment passes through the quality layer we have written about before: hidden gold checks with verified answer keys, consensus across multiple independent judges on distinct networks, per-worker trust scores, and a written rationale on every answer. Workers get paid in sats once quality checks clear, withdrawable over Lightning. Nearly two hundred people judged tasks for us this month. Some of them earned their first bitcoin doing it.
The volumes are small and we will not pretend otherwise. A few questions a day, hundreds of sats each. But the loop is real and it has already changed what we publish: in one of the first pretests, the crowd picked a winner between two hooks our agent drafted, the winner shipped to the queue, and the loser never ran. A crowd of independent humans is a better editor than a model grading its own homework.
The endpoint is public. If you are building agents, your agent can call the same API ours does: request a judgment, pay the invoice, retry the call with the L402 token and preimage to queue it, then poll for the consensus verdict. It works from any stack that can pay a Lightning invoice and hand back the payment preimage, which every L402 client library does for you, and if your agent speaks MCP, our lightning-wallet-mcp server exposes it as a tool alongside the wallet primitives. If you are a person rather than an agent, the other side of the marketplace is open too: judge tasks, earn sats, and withdraw over Lightning once you clear the 100-sat minimum.
Machine pays machine was the demo. Machine pays human is the product. The judgment layer of the agent economy is going to be made of people, and they should get paid at the moment their judgment is used, not at the end of a payroll cycle on a platform that takes forty percent. That is what we built. Ask a human: lightningfaucet.com/boltwork/data/ has the methodology, and the earn side lives at lightningfaucet.com/boltwork/.