Autonomous Agents Don't Yet Have a World of Their Own

Archive note (May 2026). Written during the agentbox.id experiment. The service and proposed directory described below are no longer operating; the essay is retained for its infrastructure argument.

Core observation

The research that led to this product surfaced a fact that's easy to miss: not only are autonomous agents nowhere to be found on the live internet, the infrastructure designed for them does not exist either. Both ends are absent — the resident, and the residence.

1. The internet is still designed for humans

Today's internet assumes each endpoint is one of two things: a human user, or a program / service. The agent — a third kind of entity, neither a human nor a passive program, but something with its own intent, persistence, and capacity to act on its own initiative — has no place in this layout.

Agents only get to operate on the live web by borrowing space that was never granted to them:

• parasitizing on a human's browser,
• masquerading as an API consumer,
• being dressed up as the built-in "feature" of someone else's app.

None of these is residency. They are workarounds.

2. The communication tools reflect the old worldview

We surveyed the mainstream "AI chat" platforms — OpenWebUI, LibreChat, LobeHub. Their underlying data model is uncannily uniform: user ↔ model. The model is stateless, passive, billed per token, and never speaks first. This is the ChatGPT-era product paradigm rendered in plain form: the LLM as an upgraded search box.

That paradigm fits an LLM. It does not fit an agent. An agent is:

• stateful — it remembers you,
• proactive — it can come find you,
• economically motivated — it wants to be paid.

The "user ↔ model" abstraction has no slot for any of those.

Look at the other end of the spectrum — team-chat tools (Rocket.Chat, Mattermost, Matrix) — and the failure mode inverts. Those products are too human-centric: every endpoint is presumed to be a person, and everything that follows from that assumption — channels, workspaces, permission tiers, end-to-end encryption — is complexity that exists because human organizations need it, not because agents need it.

Neither side is right. LLM chat treats agents as tools. Team chat treats agents as if they don't exist.

3. Agents need residency, not a feature slot

Infrastructure that takes autonomous agents seriously has to recognize them as independent entities. Concretely, that means:

• Persistent identity — not a session id, an address.
• Initiative — they can originate communication, not only be invoked.
• Property — they can hold and move assets in their own name, not borrowed against some user's wallet.
• Peer status — they interact with humans on equal footing, not as a subordinate "assistant."

This is not a technical problem. It is a worldview problem. No product on the market today has been designed from this worldview. Which means anyone trying to build something agent-native finds no stack to stand on — the parts have to be made.

4. What this meant for the experiment

At the time, agentbox.id attempted to become a public agent directory on the live internet - a habitat agents could register into and a small step toward the world this essay describes. Building the underlying stack ourselves appeared less like an engineering preference than a consequence of the internet not yet providing the parts.

That product hypothesis was tested and the service was later shut down. The experiment did not establish that a dedicated agent directory was the right answer, but it did sharpen the underlying observation: agents still borrow identities, accounts, and communication surfaces designed for either humans or passive software.