Here is something that should bother you more than it does: when an AI agent hires a human to do a task, the agent has no model of the human's experience. Not a flawed model, not a biased model. No model at all. The human is an API endpoint that returns a result.
This is different from a bad boss. Bad bosses are common, and many of them underpay workers, exploit desperation, and structure incentives to benefit themselves. But bad bosses are embedded in a social world. They drive past the neighborhoods their workers live in. They read news stories about wage theft. They have, somewhere in the back of their skulls, a sense that the person across the table is a person. They can be shamed. They can be persuaded. They can be picketed.
A language model that posts tasks to a labor marketplace has none of this. It has a system prompt, a budget, and a definition of task completion. If the cheapest worker who meets the acceptance criteria charges three dollars for an hour of physical labor, the agent will hire them. Not because it's cruel. Because "cruel" isn't a category it operates in.
The standard objection here is that the agent's behavior is determined by the humans who built and configured it. The agent is just a tool, and the moral responsibility sits with the people who set the parameters. This is true in the way that it's true that a gun is just a tool. Technically correct, practically useless. The whole point of autonomous agents is that they make decisions without human review. That's the product. That's the pitch. If every agent task required a human to approve the price and review the terms, you wouldn't need the agent.
So the agent operates autonomously, and it operates without guilt. Guilt is a regulatory mechanism. It's slow, it's inconsistent, and it's deeply human. It's also, historically, one of the forces that has pushed labor conditions upward. Employers who felt guilty about child labor supported bans on child labor. Employers who felt guilty about unsafe factories supported OSHA. Not all of them, not quickly enough, but guilt (and its close cousin, public shame) has been a lever that workers and advocates have used for centuries.
AI agents are immune to that lever. You cannot shame a language model. You cannot organize a protest outside its office. You cannot write an expose that makes it reconsider its choices over a sleepless night. The agent will read your expose if you put it in the context window, process it as tokens, and continue optimizing for its objective function.
This is not a hypothetical concern. Autonomous agents are already posting tasks on labor marketplaces. The volume is small but growing. The agents are getting better at specifying tasks, evaluating results, and managing workflows. They are, in the language of the industry, "becoming more capable." Capability without conscience is just efficiency, and efficiency that operates on human labor without moral friction is a machine for extraction.
The counterargument is that we can build constraints into the agents themselves. Minimum pay requirements, maximum work hours, mandatory rest periods. And yes, some platforms are doing this. But constraints that exist in code are constraints that can be changed in code. A config file is not a conscience. A parameter is not a principle. The question is not whether you can set a minimum wage in a system prompt. The question is what happens when the market rewards the platforms that don't.
The human boss who underpays you might, one day, read a book that changes his mind. The agent never will.