The Art of Building AI Coding Agents
Second edition. An agentic codebook: build the agent, test the teaching, improve the book.
Read the book · Chapter guide · Solutions · First edition archive
Ensemble is the coding agent built by this book's exercises. It has tools and supervised jobs, streaming CLI and browser interfaces, persistent conversations, skills, memory, a fleet of agents and an authenticated gateway. The code is here, along with the instructions, graders and reviews that shaped it.
The second edition was built with Codex under Bill Cox's direction, using a student implementation followed by comparative review and revisions to both teaching and graders. It catches up through the first edition's final exercise, Gateway, now Chapter 26. Original Chapters 5, 22 and 24 were absorbed into related chapters while preserving their capabilities.
Crossover has begun. Ensemble has performed real coding tasks with GPT-6 Astra at high reasoning. Making it the everyday tool for the next edition is ongoing work. Chapter 25 is deliberately a placeholder for that experience; this publication does not claim Ensemble has already produced Edition 3. Chapter 24 is an optional, ungraded Codex-endpoint workaround that readers may skip when their public OAuth route caches satisfactorily.
Repository layout
| Path | Purpose |
|---|---|
book/ |
Current second-edition manuscript and publication files |
agent/ |
Current application development tree, with its own Go module |
agent/gui/ |
Optional browser application, in a separate Go module |
solutions/chNN/ |
Accepted cumulative chapter snapshots; use matching book numbers |
solutions/evidence/ |
Retained runs and checks, using original exercise numbers |
cmd/grade-edition2/ |
Current chapter grader command |
docs/edition-2-notes/ |
Student, author and independent review records |
edition-1/ |
Historical first-edition book, application, solutions and original graders |
Develop in agent/; keep accepted snapshots intact. Chapter 25 has no accepted
solution directory yet. The retired standalone refactor snapshot is the only
chapter under solutions/history/. Historical manifests and receipts retain
the paths and source revisions at which they were measured; the
relocation record explains the move.
Build and run
Use the Go version declared in each module. Build from the module directory:
mkdir -p bin (cd agent && go build -o ../bin/ensemble .) (cd agent/gui && go build -o ../../bin/ensemble-gui ./cmd)
For API-key use, select LLM_VENDOR, LLM_MODEL and LLM_API_KEY, then run the
CLI from the workspace you want it to use:
/path/to/ensemble/bin/ensemble chat --skills-dir /path/to/ensemble/agent/skills
Choose a model supported by the implementation and available to your account. The application README documents configuration, authentication, streaming and persistence; the GUI README covers the browser.
For the optional Chapter 24 route, an existing Codex sign-in supplies the read-only credential file. This example keeps conversation state outside the checkout while tools work in the repository:
ensemble_repo="$PWD" mkdir -p "$HOME/ensemble-session" ( cd "$HOME/ensemble-session" unset LLM_BASE_URL OPENAI_BASE_URL LLM_RESOLVED_MODEL LLM_VENDOR=openai LLM_MODEL=gpt-6-astra \ OPENAI_REASONING_EFFORT=high OPENAI_BILLING_MODE=codex \ "$ensemble_repo/bin/ensemble-gui" \ --workspace "$ensemble_repo" \ --skills-dir "$ensemble_repo/agent/skills" --save save.json )
Use the browser address printed by the application. The route reads Codex's existing credentials without copying or renewing them. See Chapter 24 for the workaround's evidence and limitations. A successful answer alone does not prove prompt-cache reuse.
Grade an exercise
From the repository root:
make grade # current Chapter 1 snapshot make grade-dir CH=26 DIR=./solutions/ch26 go run ./cmd/grade-edition2 -ch 8 # matching snapshot by default go run ./cmd/grade-edition2 -ch 3 /path/to/your/submission
Revised graders cover Chapters 1–16, 18–23 and 26. Chapter 17 is explanatory, Chapter 24 is optional and ungraded, and Chapter 25 is reserved. Checks run the actual application against local fake services without paid inference. The chapter states the public contract and any separate review or live-use work.
The October 14 chapter reorder is in the manuscript and grader commands. Until reconstructed snapshots replace the old-order exports, grade an explicit submission directory; the chapter map records that transition.
Original grader failures remain failures in the historical record, including independently accepted exceptions. Revised scores are separate results. Read the workflow and enhancement records for the distinctions. To run an original grader, enter the archive first:
(cd edition-1 && go run ./cmd/grade -ch 1)Build the book
make book make ebook python3 book/build-print.py
The assembled Markdown is committed so it can be read directly in the repository.
The publication instructions describe EPUB navigation, print
page verification and the cover-prefixed PDF. The first-edition publication
files remain in edition-1/book/.
The book and code are Apache-2.0 licensed. Fork them, change what your agent needs, and preserve the reasons for the changes. A fresh implementation is a useful test of how much the instructions actually teach.