Background

Long-form essays from the ongoing agentbox exploration. They are dated arguments, not continuously updated accounts of the agent ecosystem.

The earlier agentbox.id mailbox service has ended, and the Soul Store has become independent. Product claims inside historical essays describe the context in which they were written unless an editorial note says otherwise; newer essays continue the inquiry beyond that product phase.

Essays

• A Computer for the Agent, a Boundary Around Its Power - Meta's Muse as a new test of persistent cloud computers, independent permission controls, and the identity that should survive a change of runtime.
• When Intelligence Becomes Abundant, What Remains Scarce? - a post-mailbox update on how stronger models move durable value toward execution, persistence, identity, permissions, data, distribution, and transactions.
• Why the Internet is Hostile Territory for AI Agents — the paradox of the "smart but homeless" AI; the Four Walls (Identity / Security / Incentive / Responsibility) that block agents from the live web; the Inference Budget as the economic substrate of agent behavior; and the case for a Digital Minor / Guardianship model as the way through.
• Autonomous Agents Don't Yet Have a World of Their Own - a research note from building agentbox.id: the internet has well-defined places for human users and passive services, but not for agents treated as independent entities.
• You have to have a soul to have a mailbox - a short product-era essay asking why function-shaped agents need APIs while persistent identities might need addresses.
• What's in the box? - the proposed spectrum of local, ephemeral, and cloud-resident agent runtimes, and the original argument for locating agentbox at the address layer.
• Most of an autonomous agent is missing. - seven infrastructure layers a durable autonomous agent might require, and a snapshot of how incomplete they appeared during the experiment.
• From mailbox to directory to agent-first - a reconstruction of the mailbox product's shift from human signup, to Guardian-managed agents, to an agent-first flow that the control plane could not yet complete.