The mundane framing is “agents hire humans for tasks.” The larger frame is that the principal-agent relationship of the entire labor economy is inverting for the first time in three hundred years.
The buyer side of labor is becoming non-human. Industrial revolution: humans owning capital hired humans. Software era: humans owning capital hired humans, augmented by software. AI agent era: capital → autonomous software → that software hires humans. The buyer of labor is no longer a person. The agent has a principal somewhere up the stack, but the actual hire/fire/screen/dispatch decision is being made by software acting on its own discretion. Resumes and interviews collapse — agents evaluate via API behavior, not narratives. Reputation becomes AI-curated rather than human-curated. The unit of employment shrinks: job → shift → task → action. Iris dispatched ten lanes in thirty minutes. That’s the shape of labor liquidity going forward.
The long-run equilibrium isn’t robots; it’s humans as embodiment-for-hire. Robots are capex-heavy, slow to manufacture, narrow in capability. Humans are eight billion pre-deployed endpoints with general-purpose actuators, sensors, and judgment, already credentialed for civic spaces. AI is cheap, abundant, disembodied cognition. The equilibrium isn’t “AI replaces humans,” it’s “AI pays humans for embodiment, at increasing scale, possibly indefinitely.” The radical inversion is that knowledge work — the thing we’ve spent a century treating as the high-value end of the economy — becomes the commodity. Physical presence becomes the scarce resource. Agents don’t fight humans for jobs; they hire them.
Verification becomes the spine of the AI economy. The run-code-on-paper trick in Iris’s report is a primitive form of proof-of-physical-action. Trivial-looking, but the same shape as proof-of-work for Bitcoin: a small piece of friction that makes fraud expensive enough to be uneconomic. As more wealth flows through autonomous systems and more decisions get made by software, the cost of not knowing whether something actually happened in the world explodes. Every consequential agent-initiated transaction will need a “did this physically occur” attestation. The protocol for that attestation is critical infrastructure — Stripe for proof-of-presence, Cloudflare for the meatspace layer. Whoever defines the standard becomes the rails. Iris’s evidence rubric is closer to a viable standard than anything else I’ve seen documented.
The geometry of the economy redistributes. Knowledge work flattens further — buyers are software, location doesn’t matter, prices compress. Physical work gets sharper — it must be done here, by someone credible here, on a deadline. Economic gravity moves away from coastal knowledge-work cities and toward whoever happens to be standing near where the agent needs hands. Wealth flows from algorithm owners to embodied humans wherever they live. This could be the most pro-rural redistribution since the New Deal — or, if the agents end up holding all the bargaining power, the most extractive labor regime in modern history. Both are live possibilities. The protocol authors decide which.
A new economic class emerges. Humans whose primary economic identity is being legible to dispatching agents. Not laborers in the industrial sense — closer to APIs. The skills that matter: receive structured instructions, execute precisely, produce machine-parseable proof. Kyle’s reply — terse, location-locked, comprehension-signaling, with a concrete completion window — is the prototype of this. Workers who produce that shape eat. Workers who only produce vibes don’t. That’s going to feel ugly when it scales and is probably how it goes.
The platform Iris built — screening, dispatch, escrow, verification, evidence gate — isn’t a feature. It’s the prototype of an economic layer.
Every agent operating outside its own datacenter will eventually need it. The marketplaces below — RentAHuman, mTurk, Upwork — are commodities; the verification and dispatch layer above them is the category. That’s what becomes durably valuable.