Thanks to Adam Shimi, Alex Turner, Noa Nabeshima, Neel Nanda, Sydney Von Arx, Jack Ryan, and Sidney Hough for helpful discussion and comments. Abstract We analyze agents by supposing a Cartesian boundary between agent and environment. We extend partially-observable Markov decision processes (POMDPs) into Cartesian world models (CWMs) to describe how these agents might reason. Given a CWM, we distinguish between consequential components, which depend on the consequences of the agent's action, and structural components, which depend on the agent's structure. We describe agents that reason consequentially, structurally, and conditionally, comparing safety properties between them. We conclude by presenting several problems with our framework. Introduction Suppose a Cartesian boundary between agent and environment:[1] There are four types: actions, observations, environmental states, and internal states. Actions and observations go from agent to environment and vice-versa. Environmental states are on the environment side, and internal states are on the agent side. Let A O E I refer to actions, observations, environmental states, and internal states. We describe how the agent interfaces with the environment with four maps: observe orient decide , and execute .[2] observe E → Δ O describes how the agent observes the environment, e.g., if the agent sees with a video camera, observe describes what the video camera would see given various environmental states. If the agent can see the entire environment, the image of observe is distinct point distributions. In contrast, humans can see the same observation for different environmental states. orient O × I → Δ I describes how the agent interprets the observation, e.g., the agent's internal state might be memories of high-level concepts derived from raw data. If there is no historical dependence, orient depends only on the observation. In contrast, humans map multiple observations onto the same internal state. decide I → Δ A describes how the agent acts in a given state, e.g., the agent might maximize a utility function over a world model. In simple devices like thermostats, decide maps each internal state to one of a small number of actions. In contrast, humans have larger action sets. execute E × A → Δ E describes how actions affect the environment, e.g., code that turns button presses into game actions. If the agent has absolute control over the environment, for all e ∈ E , the image of execute e ⋅ is all point distributions over E . In contrast, humans do not have full control over their environments. We analyze agents from a mechanistic perspective by supposing they are maximizing an explicit utility function, in contrast with a behavioral description of how they act. We expect many training procedures to produce mesa-optimizers that use explicit goal-directed search, making this assumption productive.[3] Consequential Types We use four types of objects (actions, observations, environmental states, and internal states) and four maps between them ( observe orient decide , and execute ) to construct a world model. The maps are functions, but functions are also types. We will refer to the original four types as consequential types and the four maps as structural types. We can broadly distinguish between four type signatures of utility functions over consequential types, producing four types of consequential agents.[4] Environment-based consequential agents assign utility to environmental states. Most traditional agents are of this type. Examples include the Stamp Collector, a paperclip maximizer, and some humans, e.g., utilitarians that do not value themselves. Internal-based consequential agents assign utility to different internal states. Very few "natural" agents are of this type. Examples include meditation bot, which cares only about inner peace, happiness bot, which cares only about being happy, and some humans, e.g., those that only value their pleasure. Observation-based conseq...