If an AI agent's price tag tracked its compute use instead of a wage, could its 'labor supply' scale in ways human workers never could?

This explores what happens if the price of an AI agent's work follows the compute it uses rather than a human-style wage or a flat subscription, and whether that would let agent 'labor' grow in ways human labor can't. The corpus doesn't model this question directly. It does cover the pieces, and together they suggest the answer is yes, with some surprises. The most basic piece is that an agent's supply is limited by compute, not by hours in the day. Why might AI agents bargain more patiently than humans do? argues that agents feel no cost from waiting, so they can run many negotiations in parallel for as long as they need to. If you price by compute, you are pricing the one thing that actually limits how much of the agent you can have.

The second piece is that more compute often buys better results, so a compute-linked price works partly as a quality setting. How does test-time scaling work at the agent level? finds that about 80% of the variation in multi-agent performance comes from how many tokens the system spends, not from clever coordination. How does test-time scaling work for individual research agents? shows that a research agent's search steps follow the same scaling curve as reasoning tokens. Under compute pricing, buyers wouldn't only buy more agents. They would also choose how hard each agent thinks, which is a way of scaling that hourly human work doesn't have.

Here is the twist: the token may be the wrong thing to count. Do persistent agents really cost less per token? followed a long-running agent for 115 days and found that 82.9% of its tokens were cheap cache reads. Once an agent keeps and reuses its context, the meaningful cost is per finished artifact, not per token. On the supply side, What actually drives AI data center costs? finds that servers, not electricity, make up most of a data center's cost. Compute is mostly paid for in advance as hardware, so the real limit on supply is how fast hardware gets built and paid off, not the cost of running each query. Per-call payment rails already exist: Which countries dominate agent-to-agent AI payments on chain? tracks agents paying each other for inference on-chain, mostly from US wallets.

There are two warnings. Cheap, patient, scalable agents doing the selling can also push prices up: Can independent AI pricing agents collude without communicating? shows that pricing algorithms reach above-competitive prices without ever communicating. Also, buying more compute doesn't guarantee useful work. Why do agent benchmarks not predict real economic value? finds that agents win contests but fail long professional tasks, so more compute can mean more wasted effort. The takeaway: compute pricing ties agent supply to hardware build-out and lets buyers trade spending for quality. But the cost that matters may shift from tokens to finished work, and the market structure around agents may matter as much as the cost of compute.