Chapter 1: Foundations of Agency

Part 1: The Single Agent

Agent-427 is born with read-only repository access, no durable memory, and no deployment authority. Its first task arrives before it has earned anything more dangerous. Part 1 asks how this single agent survives long enough to become useful: first through goals, environments, and actions, then through belief, representation, learning, and control. We formalize each addition and break its assumptions on purpose, so later patterns can be compared without mysticism, marketing fog, or the comforting fiction that production behaves like a theorem.

One Interpreter, Extended Four Times

We build one running interpreter: Agent-427, a program that reads a specification and carries it out step by step. Each topic adds one layer to that interpreter, one scar to its biography, and one word to our vocabulary of comparison.

Topic 1 (Goals, Environments, and Actions) gives the loop a contract. It says what counts as state, observation, action, reward, cost, constraint, and stopping.

Topic 2 (Belief, Uncertainty, and Decision) adds belief. A thermostat's sensor can drift or lie. Belief lets the interpreter act when it does not know the true state.

Topic 3 (World Models and State Abstractions) adds representation, through state abstraction. The interpreter learns to carry the useful part of its history, without dragging the whole attic downstairs.

Topic 4 (Policies, Value, and Control) adds policy, feedback, authority, and control. Single choices become governed action.

Every topic follows the same honest rhythm. First we state the assumption that makes the formalism useful. Then we violate it in the most production-like way available. Finally we ask what the interpreter must do when the lovely model and the grubby world cease speaking to each other.

Topics

1. Goals, Environments, and Actions
2. Belief, Uncertainty, and Decision
3. World Models and State Abstractions
4. Policies, Value, and Control