Edition 2026.10.2

Edition 2026.10.2 adds a second manual, a public API and MCP server for the curriculum, and fixes across the site. The curriculum stays at 523 lessons across 20 phases, with 2 lesson texts revised since Edition 2026.10.1.

Read Docker Sandboxes and Docker Agent 101

Docker Sandboxes and Docker Agent 101 explains Docker Sandboxes (sbx 0.47.0) and Docker Agent (docker-agent 1.149.0), from their purpose to each command, file, and policy object. It has 33 sections in 7 parts, 8 reference sections, and 34 figures. The PDF is attached below as aiefs-manual-docker-sandboxes-101.pdf.

• Every listing comes from a recorded run of the real binaries on one Mac. The capture kit records 219 files in three tiers: no daemon, local microVMs, and a local model through Docker Model Runner. Agent runs replay from recorded cassettes, so the check needs no model for them.
• A register of 122 conflicts between the docs, the help text, the specification, the release notes, and the captures, plus 30 rows of stale advice from older posts.
• A ledger of 22 claims about the sandbox, each with the capture that confirms it or the reason it stays unverified.
• The manuals now answer at clean URLs such as /manuals, and each manual fits a 320px phone screen.

Query the curriculum from your own tools

• Search 638 lessons and projects and read their original Markdown with /api/v1/catalog and /api/v1/resource. Errors use RFC 9457 problem details with a recovery hint.
• Connect an MCP client to https://aiengineeringfromscratch.com/mcp. The stateless server exposes two read-only tools, search_curriculum and read_resource.
• Each client gets 120 requests per minute on each function instance. The API docs cover authentication, errors, quotas, and versioning.
• Every page answers at a clean URL, and its old .html address redirects there. The homepage and the navigation pages return Markdown when a client asks for text/markdown.

Read in more languages

• Brazilian Portuguese, Persian, and Traditional Chinese join the lesson language picker (#556).
• All 33 translated READMEs are complete, including the lesson tables, and CI rejects untranslated English fragments (#558).
• On right-to-left pages, English data reads left to right, so values such as "12-780" keep their order.

Site

• Pages hold up under worst-case data: lesson names with code identifiers, counts of 1 and 1,284, empty phases, missing fields, and Persian and Vietnamese text. Counts are formatted and pluralized, long identifiers wrap, and no page scrolls sideways on a 320px phone, also at doubled text size (#566, #567).
• An empty search hides the Show more buttons, the catalog status filter lists the statuses in the data, and the glossary adds a "#" jump for terms that start with a digit.
• Phase and lesson counts on the home, about, catalog, and sponsor pages now come from the data at build time.
• The Blogs & Guides collection shows the full portfolio blogs and guides on the site, with search and format filters (#555).
• The six career routes link to shared delivery practice and a single-project portfolio guide. Lessons 14/42 and 14/54 extend the workbench capstone (#559).
• The README is reorganized around learning paths, a gradient descent preview, and runnable projects (#562).
• The sponsor page shows refreshed reach figures (#549).
• Vercel builds no longer type-check 278 lesson files inside the markdown functions, and each of those function bundles dropped from 111 MB to 28 MB (#561).

Contributors

Thank you to everyone who contributed to this edition:

• @zekierman completed the Turkish README (#526)
• @MaYangle wrote the original complete translations of 12 READMEs (#531), which #558 carries forward with authorship preserved

Sponsors

Thank you to SerpApi and NitroStack for supporting the curriculum. SerpApi provides a Web Search API for AI apps, available in Markdown and JSON. NitroStack is an end-to-end development platform for building, testing, debugging, and deploying production-ready MCP servers and applications.

Books and manuals

All six English volumes are rebuilt as EPUB and PDF files from this edition's source and attached below, with the A2A 101 and Docker Sandboxes and Docker Agent 101 manuals beside them:

Volume	Title	Chapters
1	Foundations	52
2	Deep Learning	58
3	Language	45
4	Large Language Models	68
5	Agents	157
6	Production	143
	Total	523

Interactive figures, graded quizzes, and runnable labs remain accessible through each chapter's online links. The website and README book links continue to use the latest release assets.

Full changes since Edition 2026.10.1

Edition 2026.10.1

Edition 2026.10.1 adds the first long-form manual, a catalog of 100 projects, and fixes across the site. The curriculum stays at 523 lessons across 20 phases, with 3 lesson texts revised since Edition 2026.10.

Read A2A 101, the first manual

A2A 101 explains the Agent2Agent protocol at v1.0.1, from its purpose to each request and response. It has 28 sections in 7 parts, 7 reference sections, and 31 figures. The PDF is attached below as aiefs-manual-a2a-101.pdf.

• Every request, response, and stream event in the manual comes from a recorded run of a small capture kit. The kit is three agents and a planner, in Python with no dependencies. The audit checks each listing line by line against the recorded files.
• Figures animate on the web, one step at a time, with a Replay control. Print and readers who prefer reduced motion see the final frame.
• A register of 40 conflicts between the specification, the proto, the docs, and the reference SDK, each with its upstream status. The writing found 20 more discrepancies, listed with them.
• A comparison of 14 behaviors across the Python, JavaScript, Java, Go, .NET, and Rust SDKs. The official conformance test kit was run against the capture kit, and the manual records the result.
• Every manual follows one standard, written in manuals/AUTHORING.md: Simplified Technical English limits, ranked primary sources, and figures drawn from capture files. New manuals appear on the Manuals page.

Build 48 projects, one tested stage at a time

The new Projects catalog has 100 entries across Python, Rust, TypeScript, and Go: 48 ready to build and 52 on the roadmap.

• The 48 ready projects have 196 stages. Each stage has a public input and output contract, an incomplete starter, and cumulative tests. Core tests use local fixtures and need no model credentials.
• Each ready project also has a reference implementation, mechanism figures with editable inputs, and recorded demonstrations.
• Eight new applications join the catalog: CSV Question Workbench, Document Extraction Review Desk, Calendar Focus Planner, and Feedback Theme Board. The other four are Inbox Triage Desk, Source-Grounded Study Coach, Visual Evidence Library, and Web Change Brief.
• The 52 planned projects each name a deliverable, prerequisites, a first demo, and four milestones. They are labeled as planned and make no completion or certificate claims.
• Project pages keep local progress, offer a stage-by-stage tutor, and accept complete learner reports for community certificates. Reference solutions cannot earn a certificate.

Start with one project:

python3 scripts/project_test.py semantic-notes-search --list

A2A lessons at 1.0.1

Lessons 13/19 and 16/12 taught names from before A2A 1.0, which a 1.0 peer rejects. Both now follow the 1.0.1 specification and proto. The agent card lives at /.well-known/agent-card.json, methods use PascalCase names such as SendMessage, and states are TASK_STATE_*. Each Part holds one content field, and clients send A2A-Version: 1.0. Lesson 16/03 updates its diagrams and card sample to the same names. Every JSON block and every exchange from both lesson harnesses parses against the v1.0 proto.

Site

• A Complete Lesson button at the end of each lesson marks it done. Progress bars, counters, and sidebar dots update at once, and a second control marks the lesson incomplete again.
• Color animations in 35 lesson figures now run. SMIL cannot read CSS var(), so browsers dropped 140 color animations without an error. Each now uses a static color with an opacity change, and a test fails on any new var() in a color animation.
• Beside the headline, the homepage asks whether superintelligence is here yet. Each click moves a dial and gives one of three dry replies. The card keeps its place at every width.
• Script and stylesheet URLs now come from their content. A returning visitor gets the new header and the Projects link, not a stale copy. The header keeps its compact menu at desktop widths, and command blocks have copy controls.
• The readiness checker reads its inputs as UTF-8, so it runs on Windows system code pages.
• The README shows both sponsor banners side by side on desktop, and adds star history and trending badges.

Contributors

Thank you to everyone who contributed to this edition:

• @shubhamrawat-gh added the Complete Lesson button (#482)
• @dajiaohuang made the readiness checker read its inputs as UTF-8 (#501)
• @satyampsoni updated the NitroStack sponsor banner (#546)

Sponsors

Thank you to SerpApi and NitroStack for supporting the curriculum. SerpApi provides a Web Search API for AI apps, available in Markdown and JSON. NitroStack is an end-to-end development platform for building, testing, debugging, and deploying production-ready MCP servers and applications.

Books and manuals

All six English volumes are rebuilt as EPUB and PDF files from this edition's source and attached below, with the A2A 101 manual beside them:

Volume	Title	Chapters
1	Foundations	52
2	Deep Learning	58
3	Language	45
4	Large Language Models	68
5	Agents	157
6	Production	143
	Total	523

Interactive figures, graded quizzes, and runnable labs remain accessible through each chapter's online links. The website and README book links continue to use the latest release assets.

Full changes since Edition 2026.10

Edition 2026.10

Edition 2026.10 adds exam prep for the Model Context Protocol Associate (MCPA) certification, translates the site interface around every lesson, and holds the course to stricter checks. The curriculum stays at 523 lessons across 20 phases, with 58 lesson texts revised since Edition 2026.09. CI now runs the lessons' own tests, and a sweep repaired datasets, models, tools, and links that no longer resolve.

Prepare for the MCPA exam

The new MCPA certification prep covers the Model Context Protocol Associate exam from the Agentic AI Foundation, delivered through Linux Foundation Training. It is aligned to MCP 2026-07-28, the stateless revision of the protocol.

• 34 lessons across the five exam domains, from the JSON-RPC envelope and capability negotiation to OAuth, consent, and audit trails, ending in a capstone that exercises every domain in one transcript.
• Each lesson has a runnable standard-library lab with tests, a six-question quiz, a figure, and a reference artifact.
• A 30-question diagnostic and three original 60-question mock exams, each following the blueprint's domain split.
• CI checks every lab transcript against the 2026-07-28 wire shape, so no lesson presents a legacy method such as initialize as current.

Figures in both certification programs now animate. MCPA diagrams play each message in order, and Claude labs show their inputs flowing into the result. On phones, wide diagrams scroll inside their card instead of shrinking. Certification cards now show only the exam facts each program has published.

Together, the Claude and MCPA programs hold 67 lessons, 12 assessments, and 505 original questions. Study with an AI tutor:

npx skills add rohitg00/ai-engineering-from-scratch

Then ask it to use mcpa-certification or claude-certification.

This is independent study material. It is not affiliated with the Agentic AI Foundation, the Linux Foundation, or Anthropic, does not reproduce live exam questions, and cannot guarantee a passing score.

The site interface follows your language

Lesson bodies were already translated. Now the interface around them follows the selected language too: header navigation, the lesson sidebar, "On this page", checkpoints, quiz headings, the code and output panels, and lesson navigation. It covers the eight CI languages: Simplified Chinese, Hindi, Spanish, Arabic, French, Portuguese, Turkish, and Vietnamese.

All 523 lessons have machine translations in those eight languages. Lessons revised in this edition are re-translated automatically after each merge, and code and other protected technical content stay in their original form.

Lessons that run, references that resolve

• CI now runs every lesson's own tests on each change. Lessons that need scientific libraries such as PyTorch are skipped there.
• A new check keeps quiz answers from being guessable by length. It fails if the share of questions whose correct option is clearly the longest goes up.
• References that stopped working were repaired: the Wikipedia dataset configuration, the MMAU-Pro benchmark, Depth Anything, NLI and speculative-decoding draft models, vLLM's speculative decoding and FP8 flags, GGUF conversion with llama.cpp, the current Gemini SDK and models, and GPTQModel and LLM Compressor in place of the archived AutoGPTQ and AutoAWQ.
• 46 dead links now point to moved pages, publisher records, or archived copies. Citations of pages that never existed were replaced with their real sources or removed.
• Two lesson claims now match their sources: vLLM supports speculative decoding together with chunked prefill, and the 2025 alignment-faking follow-up tested interrogation training, scratchpad length penalties, and process supervision.
• Lesson bugs fixed: MCP lessons 13/06 to 13/09 could not import their code under the documented test command; phase 19 lessons 29, 39, 59, and 86 had broken fixtures, tests, and YAML parsing; perplexity changed on every run in 10/10 (#358); the ACP audit trail in 16/03 was always empty (#357); and the few-shot code in 11/02 now matches the lesson text (#436).

Security

• Checkpoints in 19/47 load with weights_only=True, so restoring an untrusted file cannot run code.
• The code guard in 11/09 parses snippets instead of matching substrings, which closes an escape through __globals__.
• The sandbox path jail no longer lets a slashless symlink escape (#249).

Fixes around the course

• The homepage and About page load even when the browser blocks storage.
• The Docker image builds on Apple Silicon, and the setup lesson installs unzip for fnm.
• Mermaid diagrams keep readable fills in dark mode.
• Requesting Markdown at the site root now returns Markdown for agents.
• The README banner lists every phase once, and its artifact icons no longer collide with their labels.
• Smaller fixes: shared check defaults in 00/01 (#361), an unused import in 00/03 (#286), and quiz de-biasing that gives the same result on every operating system (#466).

Newsletter

Subscribe to AI Engineering Newsletter from the homepage, after any lesson, or on the About page. The form takes you to Substack with your email filled in.

Books

All six English volumes are rebuilt as EPUB and PDF files from this edition's source and attached below:

Volume	Title	Chapters
1	Foundations	52
2	Deep Learning	58
3	Language	45
4	Large Language Models	68
5	Agents	157
6	Production	143
	Total	523

Interactive figures, graded quizzes, and runnable labs remain accessible through each chapter's online links. The website and README book links continue to use the latest release assets.

Contributors

Thank you to everyone who contributed to this edition:

• @Mubashirrrr closed the sandbox path jail escape through a slashless symlink (#249)
• @thejesh23 made perplexity deterministic in 10/10 (#358) and fixed the empty ACP audit trail in 16/03 (#357)
• @FabienDaniel aligned the few-shot code in 11/02 with the lesson text (#436)
• @dajiaohuang made quiz de-biasing give the same result on every operating system (#466)
• @rushil-b-patel centralized the check defaults in 00/01 (#361)
• @ismet removed an unused import in 00/03 (#286)

Sponsors

Thank you to SerpApi for supporting the curriculum. SerpApi provides a Web Search API for AI apps, available in Markdown and JSON.

Full changes since Edition 2026.09

Edition 2026.09

The September edition brings the curriculum to 523 lessons across 20 phases, with 20 new lessons and 21 revised lesson texts since Edition 2026.08. It adds learning paths for real engineering work, expands Agent Skills and MCP practice, and improves the experience of moving between the website, repository, and terminal.

Choose a path through the curriculum

The new AI Engineering Learning Paths organize the course into four core domains, two specialist paths, and six career routes.

Explore Building and Deploying AI Applications, Software Engineering Fundamentals, Agent-Assisted Engineering, and Product Judgment and Delivery. Career routes connect existing lessons to customer AI deployment, developer experience and education, AI data systems, agent systems, LLM product engineering, and evaluation and reliability.

Each career route includes a shared foundation, role practice, a proof project, and evidence you can use to assess your readiness.

Build Agent Skills and MCP systems

• A five-lesson Agent Skills Engineering path covers portable skill contracts, discovery, invocation, permissions and sandboxes, and evaluation and packaging. Complete skill bundles include their scripts, references, and assets.
• A 17-lesson MCP path covers servers and clients, transports, resources and prompts, authorization, tool contracts, reliability, registry admission, and conformance checks, with an optional integration capstone.
• Twelve new agent-engineering lessons teach task framing, planning from evidence, isolated delegation, workflow discovery, risk, success metrics, and delivery decisions.

Install the learning skills into your coding agent:

npx skills add rohitg00/ai-engineering-from-scratch

Then ask it to use start-learning, learn-agent-skills, or learn-mcp.

A more consistent learning experience

• Clearer homepage entry points and responsive learning-path navigation.
• Theme synchronization, improved lesson and artifact cards, and corrected career-route navigation.
• Quiz prompts that keep answer keys separate and avoid revealing correct options through reply-format examples.
• More balanced placement questions and corrected answer-position bias in the reported lesson quiz.
• Updated public counts, social previews, search metadata, and machine-readable lesson access.

Translation coverage

All 523 phase lessons have published machine translations with source hashes matching this edition in Simplified Chinese, Hindi, Spanish, Arabic, French, Portuguese, Turkish, and Vietnamese. That is 4,184 current lesson translations across eight languages.

The 12 translated README landing pages are synchronized, including the sponsor section. Machine-translated lesson prose remains on the dedicated translations branch, and code and other protected technical content remain in their original form.

Certification curricula continue to be available in English through the website and repository and remain outside the translation and book pipelines.

Books

All six English volumes are rebuilt as EPUB and PDF files from this edition's source and attached below:

Volume	Title	Chapters
1	Foundations	52
2	Deep Learning	58
3	Language	45
4	Large Language Models	68
5	Agents	157
6	Production	143
	Total	523

The book renderer now preserves literal examples such as <image>, <PERSON>, and <doc A> as text in both formats, fixing malformed EPUB chapters and examples omitted from PDFs.

PDF code blocks, inline code, long table identifiers, and bare reference URLs now wrap within the page instead of being clipped at the edge.

Interactive figures, graded quizzes, and runnable labs remain accessible through each chapter's online links. The website and README book links continue to use the latest release assets.

Sponsors

Thank you to SerpApi for supporting the curriculum. SerpApi provides a Web Search API for AI apps, available in Markdown and JSON.

Full changes since Edition 2026.08

Edition 2026.08

The August edition. Three big surfaces landed: a certification prep program, the course in five languages, and a rebuilt way to navigate all 503 lessons.

Certification prep, free and build-first

A new certifications section turns public exam blueprints into study paths you build through, not read through. The first program covers all four Claude credentials from the July 2026 exam guides (Associate, Developer, Architect Foundations, Architect Professional): 33 in-depth lessons, runnable labs, diagnostics, capstones, and 295 original assessment questions.

It is AI-native: Claude Code, Codex, Cursor, or any SKILL.md agent can teach a route, run the labs, quiz you, and resume your progress.

npx skills add rohitg00/ai-engineering-from-scratch
/claude-certification

Independent community material: not affiliated with, endorsed by, or authorized by Anthropic. No live exam questions, no pass guarantees. Details in the announcement.

The course now speaks your language

Lessons are machine-translated into Simplified Chinese, Hindi, Spanish, Arabic, and Turkish, rolling out phase by phase. Pick a language from the site header or add ?lang=zh to any lesson URL; anything not yet translated falls back to English. The engine is NLLB-200 running free inside CI, with code, math, tables, and diagrams protected byte-for-byte. The README's landing copy is hand-translated into 12 languages. Details in the announcement.

Navigate 503 lessons like a map

• The roadmap is now an interactive dependency graph: see what unlocks what, top to bottom, instead of scrolling a flat list.
• The glossary grew to 243 terms across 12 learning areas.
• Every one of the 503 lessons now has an animated figure.

Learn in your terminal

Five learning skills install into any agent with one command: /start-learning builds a placement-based study plan, /learn teaches the next lesson interactively and tracks progress, /course-guide routes any topic or bug to the exact lesson, plus the placement quiz and per-phase checks.

npx skills add rohitg00/ai-engineering-from-scratch
/start-learning

Also in this edition

• Read lessons aloud with the browser speech engine, including diagram-aware narration.
• Quiz integrity fix: correct answers were biased toward the same option position across 2,026 questions; now uniformly distributed, with a CI guard.
• Statistics corrections in Phase 1: t-test p-values at large degrees of freedom and a chi-squared p-value that ran 7.5x too large.
• Homepage: cycling figure plates that explain the course, an evidence-backed learner wall, and a copy-ready install card.

Books

The six-volume EPUB and PDF editions attached below are rebuilt from this edition's sources. The book is the snapshot; the repository is the living edition.

Edition 2026.07

First book edition. EPUB and PDF for all six volumes are built by CI from the lesson sources and attached to this release automatically; allow the build a few minutes to upload them. The download links on the site and README always point at the newest release.