@Ryrenz: Alibaba Cloud turned enterprise Agent implementation into a book — ai-agent-handbook, available free in both Chinese and English. Open for just over a month, it's already racked up 1647 stars and 242 forks on GitHub. The hardest part of building Agents in the enterprise was never getting a demo to work — it's everything after that.

X AI KOLs Timeline 10/17/26, 09:05 AM News

ai-agent enterprise open-source handbook aliyun agent-architecture governance

Summary

Alibaba Cloud has open-sourced the AI Agent Handbook, an enterprise Agent implementation whitepaper available free in both Chinese and English. It covers architecture, building, running, governance, tuning, and real-world enterprise case studies, and has already gained 1647 stars and 242 forks in just over a month.

📕 Alibaba Cloud turned the story of how to implement enterprise-level Agents into a book — ai-agent-handbook, available free in both Chinese and English. Open for just over a month, it has already accumulated 1647 stars and 242 forks on GitHub — that's a fast climb. The hardest part of building Agents in the enterprise was never getting a demo to run; it's everything after it works: how to layer the architecture, how to govern it in production, who manages permissions and risk, and where to tune when performance falls short. This kind of know-how usually sits scattered across internal docs and is invisible to the outside world. ai-agent-handbook organizes it into a 30-chapter system, covering everything from architecture, building, and running to governance and tuning. The most worthwhile section is the practices chapter, with cases drawn from real scenarios in R&D productivity, design engineering, operations security and enterprise IT, customer sales and operations. The companies and products named include Geely Auto, Tastien, Chanjet, MiniMax, ShineWing, Bilibili, and PolarDB-X. The authors come from multiple Alibaba Cloud internal teams, each taking ownership of their chapter. The online site includes chapter navigation, Chinese search, zoomable diagrams, and reading position memory — it reads smoothly as a documentation site. The project team says they'll keep maintaining it and plan to add more enterprise case studies and cloud-based experiences. What paid courses often teach is exactly this — this time it's all laid out in the open. GitHub:

Original Article

View Cached Full Text

Cached at: 10/17/26, 10:55 AM

📕 Alibaba Cloud has turned its approach to implementing enterprise-grade Agents into a book — ai-agent-handbook, available for free in both Chinese and English.

In just over a month, it’s already racked up 1,647 stars and 242 forks on GitHub. That’s some serious momentum.

The hardest part of building Agents in an enterprise has never been getting a demo working — it’s everything that comes after: how to layer the architecture, how to govern in production, who owns permissions and risk, and where to turn when performance falls short. This kind of experience is normally buried in internal docs across companies and invisible from the outside. ai-agent-handbook brings it together into a 30-chapter system, covering architecture, building, running, governance, and optimization.

The most worth-reading part is the case studies, drawn from real-world scenarios in R&D productivity, design engineering, operations, security and enterprise IT, and customer sales and operations. The named companies and products include Geely, Tastien, Chanjet, MiniMax, ShineWing, Bilibili, and PolarDB-X. The contributors come from multiple teams across Alibaba Cloud, each owning their own sections. The online site offers chapter navigation, Chinese search, zoomable diagrams, and reading position memory — it works quite smoothly as a doc reader. The project team says they’ll keep maintaining it, with more enterprise cases and cloud-based hands-on experiences to come.

What paid courses often teach is exactly this kind of content — now it’s all laid out in the open.

GitHub:

aliyun/ai-agent-handbook

Source: https://github.com/aliyun/ai-agent-handbook

AI Agent Handbook

中文 | English

Organized around the Agent lifecycle — architecture, building, running, governance, and optimization — this handbook distills our practical experience from implementing enterprise-grade Agents. If this whitepaper can help even a little with individual learning or enterprise Agent adoption, it would be our greatest honor.

This project is dedicated to all the fellow travelers contributing to AI.

Read online (https://aliyun.github.io/ai-agent-handbook/) · Full table of contents · Reading guide

The reading site provides chapter navigation, Chinese search, zoomable diagrams and images, light/dark themes, and reading position memory. For local preview, use Node.js 22 to run npm ci and npm run dev. For site maintenance and GitHub Pages deployment, see the maintenance guide (https://github.com/aliyun/ai-agent-handbook/blob/main/WEBSITE.md).

1. Whitepaper Structure and Writing Background

In September 2025, we released the AI-Native Application Architecture Whitepaper (https://developer.aliyun.com/ebook/8479), covering the full DevOps lifecycle of AI-native applications — from architecture design and technology selection to engineering practices and operations optimization — systematically breaking down concepts, key points, and challenges, while attempting to offer some solutions. However, as models and agent technology have evolved rapidly, we’ve noticed that market focus has shifted from simply building agents quickly to three major new challenges:

Engineering challenge: From probabilistic intelligence to reliable productivity — Agents taking on critical tasks.
Scale challenge: Stability, security, performance, cost — from single-point experiments to intelligent infrastructure deployable at scale.
Organizational challenge: From Agent silos to intelligent organizations — Agents embedded into core business processes.

Last year’s whitepaper clearly couldn’t address these new demands.

So we restructured the whitepaper’s architecture, hoping that more up-to-date content, a higher proportion of hands-on practice, and a more community-driven collaboration model would provide a reference for enterprise technology selection and internal project approval — and that maintaining it as an open-source project would allow us to continuously present the latest thinking and implementation practices in AI-native application architecture.

2. Target Readers and What You’ll Gain

This whitepaper primarily targets enterprise-grade Agent building and adoption scenarios. It’s suitable both for engineers focused on Agent development and for internal enterprise technology selection, architecture reviews, project approvals, and cross-team alignment.

Reader	Recommended Focus	What You’ll Gain
Agent / AI application developers	Build, Run, Optimize sections	Master core engineering methods including Harness, context, state, tools, sandboxes, trajectories, and evaluation loops.
Architects and platform engineers	Architecture, Run, Governance sections	Build a complete architectural view spanning components, platform responsibilities, and the full lifecycle — and design scalable Agent infrastructure.
Technology leads and R&D managers	Architecture, Governance, Practice sections	Assess application form, maturity, investment boundaries, and production risks to support selection, project approval, and organizational alignment.
Product and business leaders	Survey report, Architecture, Practice sections	Understand which tasks suit Agents, where the boundary between human and Agent responsibilities lies, and what it takes to move from pilot to core processes.
Security, quality, and operations personnel	Run, Governance, Optimize sections	Establish observability, auditing, security authorization, pre-release validation, continuous evaluation, and failure attribution mechanisms.
Researchers and ecosystem contributors	Full book and Practice section	Learn about real-world enterprise problems, engineering abstractions, and open questions — and help build out the industry’s shared knowledge base.

After reading the full handbook, you will be able to:

Choose the minimally sufficient Agent architecture based on business goals, task determinism, and risk;
Understand the responsibility boundary between Model and Harness, and stop attributing every problem to model capability;
Design Agent task systems that can progress sustainably, be interrupted and resumed, and have verifiable completion;
Build a foundation for Agents covering execution environments, state, traffic, permissions, observability, and cost governance;
Use Traces, Trajectories, golden datasets, and evaluation experiments to create a continuous-improvement data flywheel;
Map methods to real-world scenarios like R&D productivity, design, operations, enterprise IT, and customer operations.

3. Reading Guide

Table of Contents Structure

Section	Table of Contents	Chapter Range	Focus
2026 Agent Developer Survey Report	Root directory	—	Current state of enterprise Agent development, productionization, architecture selection, toolchains, governance, and evaluation.
Preface	`00-preface/`	—	Whitepaper structure and background.
Architecture	`01-architecture/`	Chapters 1–2	Define the subject, determine the form, choose maturity, and establish a reference architecture.
Build	`02-build/`	Chapters 3–6	With Harness at the core, organize tasks, information, and actions.
Run	`03-run/`	Chapters 7–12	From stable single-Agent operation to asynchronous, multi-agent, and distributed communication.
Governance	`04-governance/`	Chapters 13–16	Make operations visible, behaviors bounded, assets manageable, and releases verifiable.
Optimization	`05-optimization/`	Chapters 17–24	Build a continuous optimization loop along two tracks: model and Agent.
Case Studies	`06-case-study/`	Chapters 25–29, Jev feature	Enterprise practices, vertical scenarios, cutting-edge Agent Infra exploration, and structured decision-making.
Conclusion and Outlook	`07-conclusion/`	Chapter 30	From Agentic Application to Agentic OS.

Chapter Navigation

Section	Chapter	Core Content
Architecture	Chapter 1: A New Stage of AI-Native Applications	Application form evolution, definition and boundaries of Agentic Applications, and enterprise maturity assessment.
Architecture	Chapter 2: Agentic Application Reference Architecture	Component view, platform responsibilities, lifecycle view, and how architecture translates from decision-making to operation and improvement.
Build	Chapter 3: Paradigm — Mainstream Harness Construction Approaches and Responsibility Boundaries	High-code frameworks, productized Harnesses, Managed Agents, Agent cloud products, and platform responsibilities.
Build	Chapter 4: Tasks — Orchestration, Long-Running Progression, and Collaborative Flow	Agent Loop, task state machines, planning, phase gates, delegation, async continuation, and completion evidence.
Build	Chapter 5: Information — Context, State, and Reusable Capability Assets	Context Builder, compression and offloading, Session, Task State, Workspace, Memory, Knowledge, and Skills.
Build	Chapter 6: Actions — Controlled Execution, Verification Feedback, and Delivery Readiness	Action Plane, Function Calling, MCP, A2A, environment contracts, permissions, HITL, and verification loops.
Run	Chapter 7: Agent Runtime and Sandboxes	Sandboxes, Agent Runtime, workspaces, environment lifecycle, and production operating conditions.
Run	Chapter 8: Agent State Storage and Semantic Assets	Event Logs, Checkpoints, workspace snapshots, Artifacts, long-term memory, RAG, and business semantics.
Run	Chapter 9: AI Gateway and Unified Traffic Governance	Identity, permissions, budgets, routing, auditing, and approval across three traffic types: LLM, MCP, and Agent.
Run	Chapter 10: Agent Asynchronous Tasks and Automation Workflows	Sync/async boundaries, task completion semantics, scheduled tasks, long-running tasks, and automated workflows.
Run	Chapter 11: Multi-Agent Collaboration and Orchestration	Heterogeneous Agent integration, team topologies, task dispatch, result aggregation, and orchestration responsibilities.
Run	Chapter 12: Agent Distributed Communication	Communication protocols and message governance for capability plane, collaboration plane, intrinsic plane, and human-machine plane.
Governance	Chapter 13: Agent Observability	Metrics, logs, Traces, events, cost attribution, and auditing.
Governance	Chapter 14: Agent Security	Prompt Injection, identity authentication, per-invocation validation, high-risk authorization, and data egress protection.
Governance	Chapter 15: Discovery and Management of AI Assets	Registration, versioning, discovery, dependency, and release management for Prompts, Skills, MCPs, and Agents.
Governance	Chapter 16: Agent Behavior Generation and Quality Verification	User simulation, environment simulation, scenario configuration, and pre-release Agent Simulation.
Optimization	Chapter 17: Model Optimization	Attribution criteria for model problems, SFT, Agentic RL, model distillation, and release acceptance.
Optimization	Chapter 18: Agent Optimization Overview	Agent optimization targets, method boundaries, and the full data flywheel landscape.
Optimization	Chapter 19: Agent Trajectory Data	From Trace to Trajectory — organizing reusable behavioral and decision evidence.
Optimization	Chapter 20: Agent Runtime Data Processing	Runtime data collection, cleaning, processing, and declarative data pipelines.
Optimization	Chapter 21: Agent Golden Datasets	Building high-quality evaluation data assets with inputs, trajectories, results, and criteria.
Optimization	Chapter 22: Agent Optimization: Bad Cases	Bad case discovery, attribution, remediation, regression, and experimental verification.
Optimization	Chapter 23: Controlled Self-Evolution	Turning effective experience into Memory, Skills, tools, and runtime mechanisms — while controlling self-evolution risk.
Optimization	Chapter 24: Agent Edge Runtime and Global Optimization	Edge runtimes, edge evaluation, performance cost, content distribution, security, and simulation.
Case Studies	Chapter 25: R&D Productivity	Code review, defect detection, patch delivery, R&D collaboration, and end-to-end delivery practices.
Case Studies	Chapter 26: Design Engineering	Vibe Designing and GenUI design paradigms and engineering practices.
Case Studies	Chapter 27: Operations, Security, and Enterprise IT	Large-scale intelligent operations practices across automotive, chain retail, and enterprise software.
Case Studies	Chapter 28: Customer, Sales, and Operations	Long-cycle memory, content insights, office productivity, and Data Agent practices.
Case Studies	Chapter 29: GOAI Agent Infra Track — Frontier Exploration of Multi-Agent Collaboration	Outstanding entries from the World AI Open Competition and multi-domain Agent Infra exploration.
Case Studies	Jev Feature: Structured Decision-Making in Practice	Jev model analysis, interface examples, and reusable decision patterns.
Conclusion and Outlook	Chapter 30: From Agentic Application to Agentic OS	From individual applications to collaborative, governable, and continuously evolving intelligent systems.

Case Study Navigation

Chapter	Case
Chapter 25: R&D Productivity	ABACI Kernel Patch Targeted Testing and Defect Detection Agent
Chapter 25: R&D Productivity	Kitta: Domain-Specific Code Review Agent
Chapter 25: R&D Productivity	PatchPilot Agents: Making Kernel Patch Delivery an Orchestrable, Verifiable Engineering Loop
Chapter 25: R&D Productivity	From Alerts to Automated Fixes: PolarDB-X’s Loop Engineering Practice
Chapter 25: R&D Productivity	From Coding Efficiency to End-to-End Delivery: Human-Agent Collaboration in Cloud Communications
Chapter 25: R&D Productivity	From Evaluation-Driven to End-to-End Delivery: AI Agent Security Product R&D Productivity Practice
Chapter 25: R&D Productivity	Multi-Agent Dev Squads: How AI-Assisted Development Moves from Writing Code to End-to-End Delivery
Chapter 26: Design Engineering	GenUI: Taking Agents from Giving Answers to Delivering Results
Chapter 26: Design Engineering	Vibe Designing: The Evolution of Intent-Driven AI Design
Chapter 27: Operations, Security, and Enterprise IT	Geely Intelligent Operations Implementation Practice
Chapter 27: Operations, Security, and Enterprise IT	Tastien’s 10,000-Store Chain Intelligent Operations Loop Practice
Chapter 27: Operations, Security, and Enterprise IT	Chanjet’s Observability and Intelligent Operations Practice
Chapter 28: Customer, Sales, and Operations	MiniMax’s Practice Building a Massive Long-Cycle Memory Data Foundation
Chapter 28: Customer, Sales, and Operations	ShineWing Accounting Firm’s Office Productivity Exploration
Chapter 28: Customer, Sales, and Operations	Bilibili’s Practice Building Full-Domain Content Insights
Chapter 28: Customer, Sales, and Operations	Operational Analytics Data Agent Practice
Jev Feature	Deep Dive into the Jev Model: How to Make Structured Decisions Quickly
Jev Feature	HiRoute: Jev Model Intelligent Routing Practice

4. Roadmap

The whitepaper will be maintained as an open project rather than being sealed after its initial release. Key focus areas ahead:

Expanding enterprise case studies: Adding more frontline cases from R&D, operations, customer service, data, security, finance, and industry verticals — covering success paths, architecture trade-offs, and real failure patterns.
Adding cloud resource hands-on experiences: Designing reproducible cloud-based experience flows around key capabilities like sandboxes, Runtime, AI Gateway, state storage, observability, and evaluation — taking readers from reading to hands-on verification.
Continuously improving governance content: Following enterprise-critical topics including identity and permissions, Prompt Injection defense, data egress, auditing, asset registration, version governance, and pre-release simulation.
Continuously improving the evaluation system: Adding methods and cases for task success rate, trajectory evaluation, LLM-as-Judge, golden datasets, bad-case regression, online experiments, and cost-quality trade-offs.
Tracking technology evolution: Continuously absorbing new advances in models, Harnesses, protocols, runtimes, multi-agent systems, and Agentic OS — and promptly correcting judgments that no longer hold.
Building community collaboration mechanisms: Gradually refining content guidelines, case templates, glossaries, proofreading processes, and version release practices to lower the barrier to high-quality contributions.

Contributions Welcome

Developers, architects, researchers, enterprise technology teams, and product practitioners are all welcome to contribute. You can:

Submit Issues pointing out factual errors, ambiguous concepts, broken links, or topics needing coverage;
Submit Pull Requests to improve chapters, correct content, optimize diagrams, or add references;
Share desensitized enterprise practices, post-mortems, evaluation methods, and architecture trade-offs;
Provide reproducible code, cloud resource experience flows, datasets, or experiment plans;
Help with terminology alignment, technical review, case review, and content translation.

Contributions should respect originality and licensing boundaries. For content involving enterprise data, customer information, internal systems, or security details, please complete necessary desensitization and obtain permission first.

5. Contributors

Thanks to all contributors who participated in architecture design, chapter writing, case compilation, and content review.

Alibaba Cloud

Contribution Area	Contributors
Preface	Ma Peng
Developer Survey Report	Ren Juan, Wang Chen
Architecture	Wang Chen, Liu Jun, Shen Lin
Build	Liu Jun, Pan Shengwei, Wang Chen
Run	Zhao Qingjie, Li Shibo, Lin Qingshan, Huang Xiaomeng, Zhang Tianyi, Zhao Yuanxiao, Sun Xiao, Song Zhen, Hu Qingda, Liu Zunfei, Zhu Tong, Yu Huafeng, Luo Xin, Kong Keqing
Governance	Xiao Changjun, Zhou Yang, Zhang Lei, Wang Fang, Zhang Haibin, Cheng Shuyi, Liu Ziming, Rao Zihao, Ren Yi, Yang Yong, Wang Shuo, Ma Xin, Liu Yuxuan, Yang Yi
Optimization	Zhang Hanmeng, Li Shengrong, Wang Yaning, Sun Jianyun, Ma Yunlei, Wang Zhen, Zheng Qianyi, Liu Hang, Chen Xin
Case Studies	Yang Tao, Zhu Yan, Yu Ailin, Hu Jun, Jiang Jian, Zhang Tianyi, Zhao Bin, Zhou Yuanhang
Conclusion and Outlook	Lin Yan

External Contributors

The project remains open to community contributions. Developers, architects, researchers, enterprise technology teams, and product practitioners are welcome to participate through Issues or Pull Requests. Once contributions are accepted, contributor information will be added to this section.

If this whitepaper helps you better understand, build, run, govern, and optimize Agents, please share it, discuss it, and join us in building it together.