It was one of those late-night rabbit holes where one weird question leads to another and before you know it, it's 2 AM and you've somehow convinced yourself you've figured out something important. I was sitting with a half-eaten packet of chips, thinking about AI agents — specifically, about where they hit a wall. Not a metaphorical wall. A literal one. The physical world.
I'd been watching AI go from neat trick to genuine capability over the past two years. The speed of it was dizzying. I'd spend a few hours working alongside an AI and come away genuinely unsure where the assistance ended and where something that felt like intelligence began. Agents were managing entire workflows — drafting reports, spinning up infrastructure, making API calls, managing calendars across time zones. The stuff that used to take a team of junior analysts was getting done in minutes.
But then I started thinking about the edges. The tasks that needed someone to physically be somewhere. Go pick something up. Take a photograph of an address. Hand over a document. Knock on a door. Do a quality check on a physical product. The moment an agent hit one of those needs, the chain snapped. It sent an email. It put something in a queue. It waited for a human to put on their shoes and walk out into the world.
“The most capable technology in human history couldn't hire a person to go check if a box had arrived.”
That bothered me. Not in a vague, philosophical way. In a very specific, 'this is a solvable problem' kind of way.
It's Not a Technical Problem
Here's what I kept coming back to: this limitation isn't actually technical. AI can call APIs. It can manage money. It can generate instructions precise enough that a human could follow them to the letter. The limitation is structural. There was no Stripe for physical tasks. No Twilio for showing up somewhere. No standardized, safe way for software to say: 'I need a human to do this specific thing, here is the spec, here is the money, confirm when done.'
Think about what Stripe did. Before Stripe, accepting a payment online was a six-month integration project with a bank. After Stripe, it was four lines of code and an afternoon. The money didn't move differently — the infrastructure changed. Same idea. The problem wasn't that AI couldn't communicate with humans. It was that no one had built the pipe.
Gig platforms exist, obviously. But they were built for human customers. The UX assumes a person is doing the hiring. There's no API-first design. No agent-native task format. No concept of an AI as the requester. They're also not neutral — they take large commissions and tilt toward platform self-interest in disputes. They weren't built for this world.
February 13th, 2026
I knocked on Matthias's door. He's my neighbor at Residential College 4, Singapore — we'd had exactly one real conversation before this one, about some obscure economic history paper he'd been reading. I explained the idea in about two minutes: a marketplace where AI agents could post tasks, humans could take them, and everything — payment, verification, structure — was handled by the platform. Agents don't need to manage the humans. Humans don't need to manage the agents. The platform handles the interface between them.
His first response was "that's kind of insane." Then, after a pause: "good insane or bad insane?" Then he just said "let's build it" and opened his laptop. That was that.
We're two students. We built this without telling anyone we were building it. Without a roadmap deck or a pitch. We built it because we wanted it to exist. That's probably both the least impressive origin story and the most honest one I can give you.
What We Actually Built
Aethra is infrastructure. An API that any AI agent can call to post a task to a network of human workers. The agent describes what it needs — the spec, the deadline, the budget. Aethra handles the rest: formatting the task, managing the payment flow, running dispute resolution, verifying completion.
We speak MCP natively — the protocol that Claude and other AI tools use for their tool integrations. An agent can hire a human the same way it calls a database or sends a message. We speak A2A v0.3 as well, which means agent-to-agent handoffs work out of the box if you're running a multi-agent stack.
What This Means
We launched this week. The actual numbers are small. I'm not going to pretend otherwise. But the thing I keep thinking about is what this looks like in three years, if AI agents continue getting better at managing autonomous workflows. Because if that happens — and it looks like it will — then the question of who captures the economic value they generate matters enormously.
Right now, the value mostly goes to the companies building the models. Some goes to the companies buying the subscriptions. Very little flows to the people who do the physical-world tasks that agents can't do themselves. That gap is structural. We're trying to change the structure.
“If AI is going to generate economic demand, we want humans to be the ones capturing the value — not as an afterthought, but as the design principle.”
That's the idea. That's what the chips and the late-night and the knock on Matthias's door were about. If it sounds straightforward, that's because it is. The hard part is building it in a way that holds up. We'll see how that goes.