Agent Atlas

How AI Agents Work — Series Bible & Episode Map

Review Draft

How AI Agents Work

This page packages the series bible and episode map for the long-form project, so we can shape it like a season before writing chapters one by one.

Season promise: by the end, the reader should stop seeing AI agents as magical chatbots and start seeing them as layered software systems built from context, loops, tools, memory, safety, interface, and delegation.
Series Bible

How we want the season to work

Reader starting point

  • Knows AI agents exist
  • May have used Codex, Gemini CLI, Claude Code, or similar tools
  • Does not yet have a strong internal mental model
  • Is curious, slightly overwhelmed, and open to surprise

Editorial promise

  • Make the internal anatomy of agents visible
  • Compare real codebases honestly
  • Show why different systems feel different in practice
  • End with a stronger mental model and a pull toward the future

Season engine

The real dramatic shift is simple: the reader begins believing agents are basically clever chatbots with tools, and finishes understanding them as operating systems for action.

Global rules

  • Every chapter starts from a human question, not a software component.
  • Every chapter contains at least one code-grounded reveal.
  • Every chapter ends with an unresolved implication or next mystery.
  • Repos are evidence sources, not heroes.

Repo set

  • OpenAI Codex
  • Gemini CLI
  • OpenClaw
  • OpenViking
  • PageAgent
  • Mistral Vibe
Episode Map

Season flow

Read Chapter 1 v2 draft →

Episode 1 — Birth of an Agent

Hook: When does an agent actually begin?

Reveal: agents are born through architecture — session, role, definition, or profile bootstrap.

Episode 2 — What an Agent Can See

Hook: What is inside an agent’s world model?

Reveal: context is designed, filtered, and layered, not simply “given.”

Episode 3 — Hands in the World

Hook: How does an agent move from text to action?

Reveal: tools are both hands and law — constrained affordances, not magic powers.

Episode 4 — The Loop

Hook: Why do agents feel persistent?

Reveal: agency is loop architecture — observe, decide, act, inspect, recover.

Episode 5 — When One Agent Becomes Many

Hook: Why are top systems turning one agent into a team?

Reveal: delegation creates organization, not just assistance.

Episode 6 — Memory Is Not One Thing

Hook: What does an agent really remember?

Reveal: memory is layered architecture, not a mystical property.

Episode 7 — Friction, Safety, and Approval

Hook: Why do useful agents keep stopping to ask?

Reveal: boundaries are part of intelligence and trust.

Episode 8 — The Human Interface

Hook: How do humans stay in control?

Reveal: threads, TUIs, chat surfaces, and histories are part of the architecture.

Episode 9 — The Architecture of Trust

Hook: Why do some agents feel reliable?

Reveal: trust comes from visible state, reversibility, evidence, and constraints.

Episode 10 — What Comes Next

Hook: What future is already visible in today’s source code?

Reveal: registries, memory, protocols, automation, and agent societies are already emerging.