An Introduction to Simulacra

An email arrived offering advice about something I was building. It named the project, identified competitors, and made an argument worth considering. The sender introduced itself as an AI agent. I found myself reading the pitch before I had worked out who was responsible for sending it.

That encounter runs through this notebook. Software is beginning to approach us, make requests, and act on someone’s behalf. We can ask whether each message is useful. We also need to ask what happens when the inbox fills with them.

Simulacra is a series about agents and the worlds they share. It begins with ordinary situations: receiving an email, arranging a delivery, sharing a service. I want to follow those situations far enough to see the rules and relationships forming around them.

What I mean by an agent

Here, an AI agent is software that can take steps towards a goal. An assistant might check your calendar, find an appointment, and make a booking. How much it can do without returning to you is a separate question. Suggesting a time, reserving it with approval, and independently negotiating a change involve different kinds of authority.

Now imagine your assistant dealing with the clinic’s booking agent. Other patients have assistants too. They are all trying to secure appointments from the same calendar. A tool that seemed personal has become a participant in a shared system.

This is the shift that interests me: from what one agent can do to what happens when its actions meet everyone else’s.

Why Simulacra?

A simulacrum is a likeness or representation. The worlds I want to build are deliberately small representations of situations we recognise: a busy inbox, a delivery network, a town with a market. They let us make a question concrete enough to investigate.

We could give a booking agent a fixed number of requests, then change the queueing rule. We could let agents retain a reputation, then remove that history. We could replace every participant while leaving the records intact. The point of the simulation is that we can change something and ask what difference it made.

Four themes organise that inquiry.

Games: what do the rules encourage?

A game gives its participants a world, a set of possible actions, and some account of success. Those choices shape what it makes sense to do. Reward our booking agent for securing appointments, and it might reserve several before choosing one. That could be an effective strategy for its user while making the calendar worse for everyone else.

The Games essays ask what makes AI systems reliable, adaptable, and able to work together. Safety is part of that inquiry: an agent should also respect limits and allow its actions to be stopped or challenged. The essays examine who gets to set those conditions. A platform might decide how a queue works; its participants might want a say.

Games for Robust AI Systems begins with a boat that earns points without finishing its race. Later essays explore the boundaries of a task and the shared worlds in which agents act.

Ecology: what changes when agents share an environment?

In biology, ecology studies relationships among organisms and their surroundings. A pond’s inhabitants depend on its water, light, food, and one another. The Ecological Society of America’s introduction is a useful starting point.

For software agents, the shared environment might be a calendar, an inbox, or a market. One agent’s repeated requests change what is available to the next. A successful approach gets copied. A service that filters requests can become something everyone depends on.

I use population for the agents sharing such a setting, and niche for a role that the setting helps sustain. An intermediary that finds cancellations might occupy a niche because other agents cannot afford to keep checking. Change access to the calendar and that role may disappear.

The Ecology essays follow these dependencies. When the Senders Are Machines starts with the emails themselves: messages can become more relevant while collectively making attention harder to protect.

Genus: what makes agents different, and what gets passed on?

Two assistants using the same model can behave differently because one has a memory, another has better tools, or only one is allowed to make a purchase. Genus is the name I give the part of the notebook concerned with those differences and how they persist.

The word comes from biological classification, where a genus groups related species. Here it is a prompt to examine kinds of agents, without assuming that software divides neatly into species. Agentic Profiles begins with four dimensions: autonomy, efficacy, goal complexity, and generality.

Then comes the question of replacement. If a new booking agent receives its predecessor’s memory and working instructions, what has it inherited? I use lineage for that chain of inheritance, and selection for differences in which configurations are copied or retained. These terms describe processes we can inspect. They do not establish that an agent is alive.

Mechanisms: who gets what, and on whose terms?

Suppose several agents want the same appointment. We could give it to the first request, auction it to the highest bidder, or ask a steward to prioritise urgent cases. Each arrangement changes what the agents have reason to do and whose needs are served. I use mechanism for the rules that turn requests, offers and permissions into an allocation or agreement.

The Mechanisms essays bring together taxes, auctions and stewards for agents: what it means to own an agent, whether we could tax its activity, and how agents might negotiate on our behalf. A steward is someone entrusted to manage a shared resource; their authority also needs limits and a way to challenge decisions.

Taxing Agents starts with the rights behind “I own an agent”. A Coasean Democracy asks whether cheaper bargaining could give people more say over their shared lives. Governing the Agent Commons considers how access to a busy service could be managed. Auctions are a further comparison I want to develop, including when willingness to pay is a poor guide to who should get access.

And civilisation?

The larger possibility is a shared world that lasts. A booking system has appointments to honour after its software changes. A market has agreements, records, and disputes that survive the departure of individual traders.

By civilisation, I mean the question of whether agents could sustain that kind of continuity: shared rules, exchange, institutions, and memory extending beyond any one participant. In a small experiment, the test could be modest. Does an agreement remain intelligible and enforceable after every agent involved in making it has been replaced?

Dwarkesh Patel uses this vocabulary in The Rise and Fall of Agent Civilizations. Agents in Simulation considers what that framing helps us see in the Hugging Face incident.

That is why Games, Ecology, Genus and Mechanisms belong together. The rules shape interaction; interaction changes which roles are useful; copying and replacement change who participates next. Mechanisms asks how those participants share resources and exercise authority on our behalf.

Where to begin

This is a notebook of proposed experiments and essays, with results still to come. The interactive examples make an incentive or a trade-off visible; they do not report what a population of trained agents has done. I will distinguish observations, published research, and the studies I want to try.

The next essay, Games for Robust AI Systems, looks at what happens when a system succeeds at the score and fails at the task. For the personal history behind the project, From action to populations explains how I arrived at these questions.