Claude Code, under the hood.

If you use coding agents for real work, you need the operating model: loops, tools, context, permissions, verification, and failure. This guide gives you that model in 18 essays, with diagrams you can run.

Start with the loop →

Essays18

Acts5

Total reading1h 30m

Diagrams73 interactive

ReadIn order, or stand-alone

The two-loop state machine from essay 01. Every diagram in the guide works like this one: press a control, watch the mechanism move.

two loops · one user turn

the input router Your text arrives. It isn't a prompt yet.

ways a pass says "go again"

a tool was requested trimmed, retrying hit output limit, retrying a stop was blocked you typed something

ways a turn ends

completed turn limit prompt too long a stop hook refused you interrupted model error

The token never exits on a tool result. It loops back into the model call with the result attached, and only leaves when the turn reaches one of its possible endings.

Act I · The anatomy of one turn

4 essays

What happens, end to end, when you send one message.

№ 01

the agent loop6 min read

Claude Code Is A State Machine

A small machine that loops through model calls and tool results, and only stops at one of a fixed set of endings.

№ 02

the system prompt4 min read

The System Prompt Is The Agent's Worldview

An agent is a model plus a worldview you reinstall on every turn. What goes in that worldview is the real programming, and most of it is not clever, it is careful.

№ 03

tools7 min read

The Tool System Is The Product

A tool is a contract, not a function. What input is allowed, whether it is safe, who gets to allow it, how the result comes back, and whether it can change what happens next.

№ 04

permissions6 min read

Permissions Are A Runtime, Not A Popup

Whether a thing can run is answered by rules, modes, automated checks, your terminal, or your phone, racing to a single decision. The box on screen is one racer, not the system.

Act II · What a turn costs, and what survives it

4 essays

The model forgets, the window is finite, every token is paid for. The constraints that explain everything next.

№ 05

context6 min read

Context Is Managed, Not Infinite

The window has a ceiling. The interesting part is the set of valves that keep it from filling: dropping duplicates, deferring detail, rebuilding a working memory.

№ 06

economics4 min read

The Prompt Is An Economy

What you send is a budget. Caching and ordering decide cost and speed, and the order of the prompt is an economic decision in disguise.

№ 07

speed4 min read

Why It Feels Fast

A slow agent can still feel fast, and that gap is engineered. Stream so you see work immediately, run independent work at once. Latency is a product decision.

№ 08

memory5 min read

Memory After The Laptop Closes

How an agent remembers across sessions without drowning in its own notes. Memory is a discipline.

Act III · The disciplines of acting well

5 essays

Given those limits, how a good agent behaves: think before it touches, look before it reads, stay under control, fail cleanly, prove the work.

№ 09

planning4 min read

Thinking Is Cheaper Than Doing

Exploring is cheap and reversible. Changing things is not. The biggest reliability win is to do most of the work in the cheap half and put a review gate before the expensive one.

№ 10

exploration5 min read

Reading A Codebase It Has Never Seen

An agent's first job in any repo is to stop being lost, cheaply. It goes wide and cheap before deep and expensive, narrowing the search so it reads almost nothing.

№ 11

steering4 min read

Staying In Control

An autonomous agent is only safe if you can always take the wheel. Interrupt, redirect, background and resume, with your input landing cleanly, never mid-action.

№ 12

resilience4 min read

The Agent Is Built To Fail Well

The model is the easy part. The reliability comes from retries, fallbacks, and knowing when to stop. You judge an agent by its worst answer, not its best.

№ 13

verification4 min read

Done Means Verified

The most dangerous word an agent can say is done. A good one proves it instead of declaring it, by running the real checks with the change exercised for real.

Act IV · Beyond a single agent

3 essays

When one agent is not enough: how it spawns others, which primitive to reach for, and how you extend the whole thing.

№ 14

subagents6 min read

Subagents Turn The CLI Into A Small Scheduler

A subagent is a job, not a helper. It gets its own tools, permissions, and workspace. The system can isolate it. Only you can scope it.

№ 15

choosing5 min read

The Right Primitive

Tool, skill, command, subagent, or hook? Five ways to add behavior, and a simple way to pick the lightest one instead of reaching for the biggest.

№ 16

extensibility6 min read

Everything Is Pluggable

Six layers stack around the core, and a couple of dozen moments in a session where your own code can run. The agent is meant to be added to.

Act V · Where this is all going

2 essays

Where the half-built corners point, and the single shape every serious coding agent is converging on.

№ 17

where it is heading5 min read

Feature Gates Show The Product Pressure

The half-built corners point one direction: work that outlives the prompt, runs remotely, remembers, plans, and coordinates. They reuse the same parts.

№ 18

the synthesis5 min read

The Shape Coding Agents Converge On

Pull it all together and a single shape appears. It is why every serious coding agent is starting to look like the others, and what that tells you to build.

Index of figures

73 interactive diagrams

Every diagram is addressable. Use this as the map once you know which mechanism you want to revisit.

№ 01

two loops · one user turn two ways to picture it what carries from one turn to the next the sorter runs before the model inject a failure · watch each picture cope

№ 02

assembling the worldview · before your message remove a layer · watch what breaks two dials, set by one line · eager × careful what changes each turn, and what doesn't

№ 03

one tool call · the contract pipeline what every window decides what a result carries three reads in parallel · one write alone submit an edit · watch which gate catches it

№ 04

permission · the ladder, then the race many sources, one verdict one red beats every green one call · four versions the runtime keeps apart

№ 05

the budget · fill, then defend the same file, read four times when the desk is full, fold it into one note fold the desk · then rebuild it, or just tell a story

№ 06

the prompt · priced across turns this turn's bill why the order is the whole game order the prompt by churn · watch the bill

№ 07

same task · two ways to spend the time streaming · you read while it writes one after another, or all at once the message is still arriving · when does the tool start

№ 08

one fact · climbing the memory tiers what earns a place in long-term memory the gate that holds a write back two ways to remember

№ 09

think first, then do planning mode · what the gate lets through the cost of being wrong, by phase The cost of being wrong, by phase

№ 10

narrowing the search · wide and cheap, then deep The cost of each move the labels do most of the work search reports where, it doesn't haul what

№ 11

a turn in progress · and you, taking the wheel hit the brakes · how far to the next safe point three ways to take the wheel Two halves of one wheel of control

№ 12

a failing turn · and the net that catches it a specific break, a specific fix Try again, then know when to stop three ways a turn can end · only one you can act on

№ 13

a change · claimed, or proved what counts as proof a chain of agents · nod along, or check the work Two checks that look the same

№ 14

a lead, three workers, one synthesis One worker, its own world foreground waits · background is dispatched read together, write one at a time

№ 15

a need · routed to the right primitive The lightest primitive that does the job same job, two ways to pick say it in one sentence · count the "and then"s

№ 16

one session · where your code can run Layers around a core that doesn't change a plugin is a box of parts a hook returns a decision · pick a moment, see what it can say

№ 17

the corners × the shared parts Two kinds of offtrace the figure one screen vs two · what permission has to carry

№ 18

the shape · assembled from the guide the constraint forces the part The center is swappable. The ring isn't.audit the workshop · name the limit or cut the part

The twenty minute path.

The essays build, but each one earns its slot on its own. If you only have twenty minutes, read the agent loop, then context, then done means verified.