zero2Agent: Zero-to-Production Agent Engineering Tutorials
A systematic Agent engineering tutorial site for programmers, covering core mechanics, framework teardowns, hands-on projects, and deep dives into big-tech interview questions.
This repository is fundamentally a documentation/tutorial site with no executable Agent runtime, giving a small attack surface: only two GitHub Actions (Pages deploy uses appropriately minimal permissions; the daily star-history workflow runs a self-forked action via mutable tag @v2 rather than a SHA with contents:write, a supply-chain concern — hence deductions on least_privilege/dependency_security). Sensitive-data handling has positive evidence in skill tests (safe_summary excludes the token). Source attribution is excellent: THIRD_PARTY_NOTICES, pinned commit baselines, itemized acknowledgments — full marks. Rollback has no recovery mechanism or notes, scored 0. user_confirmation has no counterpart in a pure docs product; scored 1 on existing evidence only.
README numbers and module status are internally consistent (131 articles, 12 complete + 1 in progress), but consistency is only ordinary: scripts lack a unified documented entry point. Dependency availability is weak: only 'bundle install' is given with no lockfile evidence; skill scripts need LLM API configuration whose user-side docs are absent from the sample. failure_messages rests only on test assertions, with no user-facing error documentation — scored 1.
Audience definition is the strongest area: explicit 'who this is / is not for' lists and well-covered scenarios (job hunting, career switch, deployment) — scored 3. Capability boundaries are partially declared (in-progress modules, community material not treated as API spec) but timeliness risks are not systematically stated — 2. trigger_precision is 2 based on the out_of_scope/review gating logic visible in skill tests; environment_fit is 2 with both local Jekyll and Pages paths.
Information architecture is excellent: 13 modules, directory tree, navigation, consistent tabular overviews. Examples and interview Q&A structure (novice vs expert answers, frequency ordering) earn 3. License is 1: README/LICENSE say MIT but the green-paper PDF declares CC BY-NC-SA 4.0; the dual-licensing relationship is not clarified at the LICENSE level and the 2026 copyright year is questionable. versioning_changelog is 1: only 'PDF v1.2.0', no CHANGELOG. known_limitations partially covered via in-progress markers, 2. install_notes complete but lacking dependency locking and troubleshooting, 2.
As a tutorial product, outputs (online site + PDF + companion repo) are directly usable — 2. Marginal value is high: 131 articles, 13-framework comparison, real interview-frequency analysis differentiate it clearly — 3. cost_benefit 2: low reader cost, but accuracy depends on manual verification with no automated validation evidence.
Key factual baselines are pinned to specific commits (pi-mono a470b121, deepseek-harness rc.8, ai-agent-book SHA) — good traceability, 2. Cross-source corroboration 2 via multiple independent repos and official docs, but quantitative claims like '630 questions' and 'real big-tech questions' are not statically verifiable. fact_inference_separation has local evidence (explicitly separating official API spec from community material), 2 not 3 because no systematic full-text labeling is shown.
- star-history.yml references a self-forked action via a mutable tag with contents:write; pin to a commit SHA and narrow permissions.
- The MIT license and the green paper's CC BY-NC-SA 4.0 dual-licensing boundary is unclear; confirm applicable terms before redistribution or commercial use.
- Quantitative claims such as '630 real interview questions' cannot be verified in static review; spot-check before relying on them.
- No rollback or versioning mechanism; content freshness depends on the author. Read alongside the pinned commit baselines (Pi, DeepSeek Harness).
What does this agent do, and when should you use it?
zero2Agent (GitHub: ranxi2001/zero2Agent) is a Jekyll-built tutorial website for Agent engineering, published via GitHub Pages at onefly.top/zero2Agent. It targets developers who can code and use AI tools but have never built Agent systems systematically, as well as new grads and career changers preparing for AI Agent job interviews. The site contains 131 articles across 13 modules: Agent Basics (Tool Calling, Memory, Context, Loop), LangGraph, three first-party SDKs, surveys of 13 mainstream frameworks, OpenClaw, Claude Code, DeepSeek Harness, Pi Coding Agent, Codex CLI, Agent training and application practice, interview prep, and a Final Project (OfferPilot) built on a hand-written 10-layer Harness. It also ships a '500 Agent Interview Questions' green-book PDF (v1.2.0) with 630 high-frequency questions sourced from Ant, Alibaba, ByteDance, Tencent and other major tech companies. The code repository is MIT-licensed; the PDF is CC BY-NC-SA 4.0. It is a learning and interview-prep resource, not a runnable Agent product.
zero2Agent is delivered as a static tutorial site. Markdown articles live in per-module directories — learn-agent-basic, learn-langgraph, learn-sdk-frameworks, learn-agent-survey, learn-openclaw, learn-claude-code, learn-deepseek-harness, learn-agent-training, learn-agent-practice, learn-pi, learn-codex, learn-agent-interview, final-project — and are built by Jekyll (_layouts, _data/nav.yml, assets/docs.css, assets/app.js) into a site with a three-column layout, sidebar, TOC and Mermaid diagrams. Content-wise: Agent Basic covers model APIs, Tool Calling, Memory patterns and context compression; the LangGraph module demonstrates StateGraph, add_conditional_edges, Fan-out/Fan-in and Prompt Chaining; the SDK module compares OpenAI Agents SDK, Google genai SDK and Claude Anthropic SDK; the survey module covers 13 frameworks including AgentScope, Mastra, Semantic Kernel, Eino, DeerFlow, LangChain, Google ADK, AutoGen and Vercel AI SDK; the Claude Code module builds a full Coding Agent from a 30-line loop over 12 lessons (Loop, Tool Dispatch, Subagent, Agent Teams, Worktree Isolation); the Final Project tears down OfferPilot (companion repo ranxi2001/OfferPilot), a hand-written 10-layer Harness with a Query Engine, an FTS5/embedding dual-channel knowledge base, a sub-agent concurrency pool, and Whisper/FunASR speech diagnostics. examples/agent-api-lab provides key-free API protocol, context-ablation and fault-injection experiments. The interview module compares 'novice answers vs expert answers' across 17 assessment dimensions (architecture selection, tool management, memory and context, multi-agent collaboration, Agent Infra, and more).
- A Python developer who has never built LLM applications and wants to learn Agent core mechanics (Tool Calling, Memory, Context, Loop) rather than just running demos
- A job seeker preparing for AI Agent roles at Ant, Alibaba, ByteDance, Tencent or Ctrip, using the interview module and the 500-question green book for rapid review across 17 dimensions
- An engineer who uses Claude Code or Cursor but wants to understand the internals, following 12 lessons to hand-write a Coding Agent with Subagents, Agent Teams and Worktree isolation
- A team choosing among LangGraph, OpenAI Agents SDK, Google genai SDK and Claude Anthropic SDK, using the three-SDK comparison and 13-framework survey
- A developer who wants to ship an Agent to production, following the Final Project's OfferPilot path from PRD through the 10-layer architecture to three deployment options
- An engineer exploring Agent training (SFT, RL, GRPO/PPO, data mixing, eval feedback loops) via the training-practice module
What are this agent's strengths and limitations?
- Complete curriculum: 131 articles across 13 modules forming a closed loop from fundamentals and framework teardowns to production deployment and interview prep, with per-module completion status clearly labeled
- Scarce interview material: 630 questions across 17 assessment dimensions drawn from real interview experiences at Ant, Alibaba, ByteDance, Tencent, Ctrip and Baidu, preserving novice-vs-expert answer comparisons
- Beyond concepts: the Claude Code module hand-writes a Coding Agent over 12 lessons, and the Final Project implements a 10-layer Harness without LangChain/LangGraph, with reproducible steps, failure experiments and acceptance evidence
- Transparent fact baselines: the Pi module is anchored to official commit a470b121, the DeepSeek Harness module pins API dsh-v0.1.0-rc.8, and THIRD_PARTY_NOTICES.md discloses all sources
- The Agent Application Practice module (5 articles) is still in progress, so that track is incomplete
- The repository's core deliverable is a tutorial, not a runnable Agent product — it cannot be deployed as a service, and capabilities must be implemented by the reader
- It depends heavily on third-party projects (Claude Code course, pi-mono, DeepSeek Harness, etc.); where community materials are based on older versions they may diverge from current APIs, requiring checks against THIRD_PARTY_NOTICES.md
- The green-book PDF is CC BY-NC-SA 4.0 (non-commercial, attribution required); local builds need a Ruby/Jekyll environment with no documented version requirements
How do you install or deploy this agent?
Running locally requires Ruby and Jekyll:
bash
git clone https://github.com/ranxi2001/zero2Agent
cd zero2Agent# Install Jekyll (requires Ruby)
gem install bundler jekyll
bundle install# Local preview
bundle exec jekyll serve
# Visit http://localhost:4000/zero2AgentNo API keys are needed to read the tutorials; examples/agent-api-lab specifically provides key-free API protocol, context-ablation and fault-injection experiments. For online reading, simply visit https://onefly.top/zero2Agent — no installation required.
Missing information: no pinned Gemfile version list or minimum Ruby version is documented.
How do you use this agent?
Two main usage paths:
- Read online: visit https://onefly.top/zero2Agent and follow the learning path — Agent Basic → LangGraph → SDK Frameworks → Framework Survey → OpenClaw / Claude Code / DeepSeek Harness / Pi / Codex CLI → Training & Application Practice → Interview Prep → Final Project. Before interviews, download publish-pdf/output/zero2Agent-绿皮书-Agent面试500问.pdf (630 questions, 17 dimensions, ordered by real-world interview frequency within each topic).
- Build locally: clone the repo and run bundle exec jekyll serve, then browse at http://localhost:4000/zero2Agent. Work through code examples against the referenced repos (learn-claude-code, Learn-OpenClaw, pi-mono, OfferPilot, etc.).
How does this agent compare with similar options?
The framework-survey module directly compares 13 mainstream Agent frameworks, including LangChain (LCEL, RAG, and when not to use it), LangGraph (StateGraph graph structure), AutoGen and AgentScope (multi-agent conversation and distributed collaboration), Mastra (TypeScript-native), Semantic Kernel (Microsoft plugin ecosystem), Eino (ByteDance, Go, high concurrency), DeerFlow (Deep Research), Google ADK and Vercel AI SDK. The Claude Code, Pi and Codex CLI modules additionally compare mainstream coding agents' architectures. Readers can use it as a Chinese-language pre-selection research survey.