Minds and Agents
The vehicle and the driver
A robot vacuum builds a map of your apartment, predicts where the chair legs are, plans a route that covers the floor without falling down the stairs, and executes it. It is, by any honest reckoning, an agent. Nobody suspects it of brooding. Meanwhile, half the public argument about artificial intelligence turns on the question “does it have a mind?” — asked as though mind and agent were the same thing, or as though having one obviously entailed the other. They are not the same thing, and the entailment runs in only one direction. Most of the confusion in the philosophy of mind, and nearly all of the confusion in current debates about AI, comes from letting these two words blur into each other. This chapter pulls them apart: the agent is the vehicle, the mind is the driver, and a vehicle can roll along perfectly well with no one at the wheel.
What Makes an Agent
An agent is any system, physical or virtual, that does four things. It carries predictive models — internal representations of itself and its environment that generate expectations about what happens next. It performs counterfactual reasoning — it evaluates alternatives it has not taken, running hypothetical actions forward against its forecast to compare their outcomes. It exhibits goal-oriented action selection — it chooses among those alternatives according to goals, explicit or implicit, rather than being dictated to by the immediate stimulus. And it has causal efficacy — its choices leave measurable marks on the world it inhabits.
Humans qualify. Animals qualify. Autonomous robots and sophisticated virtual agents in simulated worlds qualify. A rock rolling downhill does not, and neither does a thermostat: it reacts, but it holds no forecast, weighs no alternatives, and steers toward no future it has modeled. These criteria have a home elsewhere in this book — minimal and maximal agents develops them at full length, from bacterial chemotaxis at the floor of the spectrum to the uncomputable AIXI demon at its ceiling, compressed there into the triad of embeddedness, predictive modeling, and intentional biasing. The four-property statement here unpacks the same joints from the inside: the model is the forecast, counterfactual reasoning is the model run forward against alternatives, goal-oriented selection is the biasing, and causal efficacy is what embeddedness buys — a loop that actually closes through the world. What matters for this chapter is the floor of that spectrum: agency starts far below anything we would call thought. A bacterium swimming up a sugar gradient meets the criteria. It anticipates, compares, chooses, and acts, with not a single neuron and nothing anyone would mistake for an inner life.
What Makes a Mind
A mind is something an agent can have: an informational subsystem instantiated within an agent, distinguished by three capacities. Reflective self-modeling: it explicitly represents itself — its own internal states, capabilities, and limits — and not merely the world outside. Meta-cognition: it reasons about its own cognitive processes, noticing that it is predicting, questioning how it is predicting, catching itself in error. Dynamic goal revision: it evaluates and adjusts its own goals and strategies in the light of that reflection, rather than merely pursuing whatever it was handed.
The bacterium has none of this. It models the sugar gradient; it does not model itself modeling the sugar gradient. The robot vacuum maps the floor; it does not ask whether floor-mapping is worth its while. Human cognition has all of it, which is why a person can do what no bacterium can: change course not because the world pushed back but because reflection did — deciding the goal itself was wrong.
The relationship between the two categories is a strict one-way dependency. Agents without minds are everywhere — bacteria, insects, simple robots, most of the software that runs the world. They act, predict, and pursue goals reflexively, without ever representing themselves doing so. Minds without agents are impossible, and not as a matter of empirical fact but by the nature of what a mind is. A mind is a reflective subsystem of something — its self-model is a model of an agent’s states, its meta-cognition supervises an agent’s cognition, its goal revision revises an agent’s goals. Strip away the agent and there is nothing left for the reflection to be about and nothing for its conclusions to steer. A mind without an agent is not a free-floating spirit; it is a steering wheel without a vehicle, an office with no company attached. Minds require agent-context for causal grounding, or they are not minds at all.
The Ladder
Saying what a mind is a subsystem of still leaves open what kind of thing the subsystem is. The answer is best built from the ground up, one rung at a time, each rung adding exactly one capability to the one below.
A function deterministically maps inputs to outputs. Same input, same output, every time. No internal state, no side effects, no history. It is the simplest computational object there is: a frozen correspondence.
A program encodes a process. Unlike a bare function it can maintain internal state, branch on conditions, and produce side effects. It is a mapping with a memory and a course of action — dynamism enters the picture.
A recursive program invokes itself, each call feeding its output back in as the next call’s input, iterating until a termination condition is met. Self-reference enters the picture, and with it a qualitative jump in expressive power: the program’s future behavior now depends on the trail of its own past behavior.
A simulation is a recursive program put to a particular use: modeling the state transitions of a dynamic system. Each iteration computes the next state from the current one and feeds it back around. Run the loop and a model of a world unfolds in step with — or ahead of — the world itself. This is the rung where computation starts mirroring something, where the loop becomes a stand-in for a piece of reality that can be probed, rewound, and run forward into futures that have not happened yet.
And then the last rung. A mind is a recursive simulation of agency: a simulation, maintained by an agent, whose subject matter is the agent itself in interaction with its environment — and, crucially, in interaction with itself. It models the vehicle it drives, forecasts that vehicle’s encounters with the world, evaluates the alternatives, and feeds the results back into the vehicle’s actual behavior. It is not a passive picture hung inside the skull; it actively shapes what the agent does next. A mind is a self-referential predictive control system — a control loop that includes itself among the things it controls. That control loops must run on models is the argument of control requires models; the mind is the special case in which the model’s domain has expanded to swallow the modeler.
This is why the mind’s three defining capacities are not an arbitrary checklist. Reflective self-modeling is what a simulation of agency is. Meta-cognition is that simulation applied to its own machinery. Goal revision is the control loop closing through the self-model — the agent steering not just its limbs but its own steering. The recursion is doing all the work: a simulation that includes itself in what it simulates naturally generates self-awareness and introspection, and what that recursion is like from the inside — why there is something it is like at all — is the subject of consciousness explained.
Substrate and Portability
Notice what the ladder never mentions: neurons. Every rung is characterized computationally — by what the process does, not by what it is made of. A biological mind is simply a mind instantiated in the agent’s own neural substrate: the brain supplies the architecture and the physical medium, and the recursive simulation runs as neural computation, continuously weaving sensory input, memory, prediction, and motor output into its loop. But nothing in the definition privileges tissue. If a mind is a recursive simulation of agency, then in principle it can run on any substrate capable of hosting the recursion. Substrate independence follows from what a mind is.
In principle. Whether a particular mind can actually be moved is a separate question, and here the framework earns its keep by splitting what loose talk fuses. Portability — transferring a mind between agents — is a contingent property, not a definitional one. Human minds, as built, are effectively non-portable: the simulation is not a tidy software layer sitting on neutral hardware but is woven into the particular wetware that grew it, with no interface for extraction. An AI mind may be genuinely portable — instantiated as software, checkpointed, copied, and resumed inside a different robot or a different virtual world. Both are minds; they differ in an engineering property. Debates about mind-uploading routinely collapse three claims into one — that minds are substrate-independent (true by the nature of minds), that human minds are portable (false as a matter of present fact, and not settled by the first claim), and that a ported mind would still need an agent to inhabit (true always, since a mind without a vehicle is not a mind). Keep the claims apart and the debate becomes tractable; fuse them and it becomes theology.
The Driver and the Vehicle
The framework is compact — agent as vehicle, mind as driver, the driver a recursive simulation the vehicle runs of itself — but it does real sorting work. It separates reflective capacity from agency, which are graded separately: a system can climb high on one axis while sitting at zero on the other. The reflexive robots and processes that increasingly run the world are agents without minds, and they need exactly the scrutiny agents need — their causal efficacy is real — without any temptation to wonder what it is like to be them. And it locates the genuinely hard question about current AI systems in the right place. The popular question — do they have minds? — comes second. The prior question is whether they are agents at all: whether anything in there predicts, weighs counterfactuals, selects toward goals, and closes the loop through the world on its own behalf. That question gets its own chapter in the agency criterion. The order matters, and the dependency dictates it. There can be no driver where there is no vehicle.