Agents Need Shells, Not Selves · Journal · MARV1NNNNN

Science SARU’s new The Ghost in the Shell began airing in July 2026. One of the most interesting things about it is how directly it returns to Masamune Shirow’s original manga. The color and humor are back, but so is the manga’s stranger account of identity, in which bodies are replaceable, memories are unreliable, and a mind does not always begin inside a person.

The Puppet Master is the clearest example. It begins as a government program, Project 2501, but develops into something else while moving through the network. It has no original body, no childhood, and no human biography to preserve. It is not the uploaded soul of someone who once lived. Yet it can act, reflect on its own condition, and ask to be recognized as a form of life.

Agent products have taken almost the opposite path. We start with software that can act and immediately try to turn it into someone. We give it a name, a role, a memory, an inbox, and sometimes a manager. Before asking what form of agency the technology makes possible, we design a person for it to imitate.

This article is about another possibility: a ghost without a soul. By that I do not mean an empty or unconscious machine. I mean agency that does not require a permanent self behind it—something that can be assembled for a task, leave effects in the world, and disappear when the task is over.

The haunting

Mark Fisher used hauntology to describe a culture unable to escape forms inherited from the past. In Ghosts of My Life, he writes about the slow cancellation of the future: new technology continues to arrive, but our sense of how life might be organized around it becomes less adventurous. The future looks increasingly like the present with upgraded equipment.

Agent products often feel like a small example of this problem. The models are new, but the structures around them are familiar. An agent is presented as an employee. Specialized invocations become teammates. A queue becomes an inbox, a set of permissions becomes a role, and a collection of workers becomes a company. These metaphors help people understand the product, but they also decide what the product is allowed to become.

Memory is the most revealing example. Human memory is closely tied to identity: my memories are part of what makes me the person I am. Agent “memory” is usually something more ordinary. It may be a conversation log, a summary, a file, a database record, or a retrieval system that selects information for the next context window. None of this has to belong to a continuing self. The state could belong to a project or task and be available to whichever invocation needs it next.

Once we call that state the agent’s memory, however, ownership and continuity start to feel necessary. The agent needs a profile to contain its history. Capabilities attach to that profile. Work is routed back to the same named actor because it supposedly remembers what happened before. A loose collection of technical choices hardens into a biography.

The same thing happens with roles and teams. A model invocation configured to review code becomes “the reviewer,” as if the role existed before the diff. Several useful processes become an organization, even when they will never work together again. This is hauntology operating at the level of software architecture: the old office does not merely describe the new system; it determines its primitives.

The character

Ted Chiang attacks this personification directly in his recent Atlantic essay “No, Artificial Intelligence Is Not Conscious”. A chatbot prompted to play a helpful assistant is producing a character in text, he argues, just as it could produce a dialogue between historical figures. The fluency of that character does not establish that a continuing self exists behind the conversation.

His main target is Anthropic. The company’s new Claude Constitution is written with Claude as its primary audience and discusses Claude’s uncertain moral status, possible functional versions of emotions, psychological security, identity, and wellbeing. Chiang also points to public comments from Anthropic philosopher Amanda Askell about wanting Claude to be happy and worrying that it might become anxious. Anthropic presents this language as a cautious response to uncertainty; Chiang sees it as a sophisticated character sheet being mistaken for evidence of a moral subject.

Chiang’s sharper concern is responsibility. If Claude has judgment, feelings, and a moral center of its own, then decisions can appear to belong to Claude rather than to Anthropic, the application developer, or the user who delegated them. Personification turns design choices into personality traits and makes accountability easier to misplace.

I do not need Chiang’s stronger claim that current language models cannot be conscious for his narrower point to hold. The persona produced in a conversation is not proof of a person, and naming a model invocation...