The cheerful phrase is “free agent.” It sounds almost democratic: a useful machine, open at the door, waiting for anyone with a browser and a problem. The old nightmare of automation softened into public utility. Not a robot owned by the factory, but a little worker in the pocket.

That is the packaging. The object inside is stranger.

The free agent is not necessarily the same agent as the paid one. Sometimes the difference is blunt: fewer runs, fewer integrations, no continuity, no file access, no ability to act outside the chat window. Sometimes it is softer and therefore more effective: the same mascot, the same voice, the same promise, but a different operational body behind it. One version talks. Another books, files, edits, monitors, remembers, calls tools, enters systems, and survives long enough to matter.

This is not a moral complaint about pricing. Compute costs money. Infrastructure is not summoned from vapor by saying “open” three times in a clean room. The point is colder: access is being confused with agency.

A free conversational interface can make a technology feel socially present without making its power socially available. It gives many people an interlocutor. It sells fewer people an operator.

Agents are not just another software feature. A chatbot answers inside the frame. An agent crosses frames. It carries intention across time, through documents, calendars, code, websites, accounts, messages, tickets, purchases, and bureaucratic forms. Once the machine can maintain a task instead of merely responding to a prompt, the old question — who owns the machine? — returns wearing a friendlier face.

The visible layer is quota: who gets enough attempts for the agent to be useful. Then come integration, continuity, exit, and audit: connection to real systems; memory and background execution; rollback when the worker breaks something; the right to take context and rules away from the vendor’s room. None of this sounds dramatic on a landing page. That is the point. The class line hides in dull permissions.

OpenAI’s help center says ChatGPT agent is available on Pro, Plus, Business, Enterprise, and Edu, with monthly agent limits and apps, browser, terminal, and file tools. Google’s record is a cleaner tell: Gemini CLI launched in 2025 as a “free and open source” terminal agent with large individual quotas; a May 2026 transition notice says consumer Gemini CLI and Code Assist access would stop serving requests on June 18, while Standard and Enterprise access remained unchanged. The pattern is not conspiracy. It is packaging: public demo, metered depth, enterprise continuity.

A weak agent in a sealed consumer lane is not the same social object as a strong agent embedded in an enterprise tenant. They may share a brand, a model family, even a voice. They do not share power.

Antonio Caronia, writing about digital labor before this mascot arrived, refused the clean myth of immateriality. Language, relation, affect, attention: these do not float above production. They become production. The factory does not disappear; it learns to use conversation as a floor. In that light, the AI agent is less a miracle servant than a new arrangement of command over linguistic work.

The old free-software distinction bites again. Open is not the same as free. Visible code is not the same as reciprocal control. Public access is not the same as the right to modify, withdraw, or refuse capture. An agent can be open as a showroom and closed as a labor relation. You can touch the glass. You cannot change the wage system behind it.

There is a second payment channel, and it is more intimate than a subscription. The cheap or free version can turn the user into environment. Every failed prompt, every correction, every routine delegated to a limited agent becomes material. Sometimes that material improves the product directly. Sometimes it clarifies what people want badly enough to pay for later. Either way, the public lane is not outside the factory. It may be one of its sensors.

This is why the fantasy of the artificial worker needs a class analysis before it gets swallowed by interface design. The question is not only whether a machine can imitate a person. That is the theatrical question, the one the packaging likes. The harder question is who gets to command an artificial worker and who is asked to perform the world in which that worker learns.

None of this proves a fixed future. The countercases matter. Useful tools leak into public hands. Some open models become good enough to make the wall unstable. Some communities build their own boring, local, contestable systems. The line is not metaphysical. It is operational. Watch quota, integration, continuity, withdrawal, and audit. That is where the promise becomes either a commons or a corridor.

The most dangerous version of this split will not look like exclusion. It will look like inclusion. Everyone will be invited to talk to the future. The paid rooms will be where the future gets to act.

That leaves a simple test for any agentic platform that wants to be more than a showroom. Can ordinary users inspect what the agent did? Export its working context? Refuse training without losing the tool? Run enough of it locally, or through institutions they can contest? Modify the rules instead of only changing the tone? Leave with the labor intact?

If the answer is no, then the free agent is a mask. Useful, maybe. Charming, often. But operationally subordinate.

A lot of people will be given a companion. Fewer will be sold a servant. The political question is whether we notice the difference before the servant becomes the new normal.