1.8K ·GitHub
What is this?
An open-source Alibaba Cloud whitepaper (handbook) distilling hands-on experience building enterprise-grade AI agents across their full lifecycle: architecture, building, running, governance, and tuning. It is maintained as a living open project rather than a one-off release.
An open-source Alibaba Cloud whitepaper (handbook) distilling hands-on experience building enterprise-grade AI agents across their full lifecycle: architecture, building, running, governance, and tuning. It is maintained as a living open project rather than a one-off release.Is it a book I read, or software I install and run?A book, not software: about 30 chapters of markdown grouped into sections — Architecture (1-2), Build (3-6), Run (7-12), Governance (13-16), Tuning (17-24), Case Studies (25-29) — stored in this repo. There is nothing to install or run.
A book, not software: about 30 chapters of markdown grouped into sections — Architecture (1-2), Build (3-6), Run (7-12), Governance (13-16), Tuning (17-24), Case Studies (25-29) — stored in this repo. There is nothing to install or run.Where should I start if I'm new to enterprise agents?The recommended first-time path is the 2026 Agent Developer Survey Report, then chapters 1-2 (architecture), 3-6 (building), and 13-16 (governance). The README also lists paths for production rollout, multi-agent systems, evaluation work, and vendor selection.
The recommended first-time path is the 2026 Agent Developer Survey Report, then chapters 1-2 (architecture), 3-6 (building), and 13-16 (governance). The README also lists paths for production rollout, multi-agent systems, evaluation work, and vendor selection.Is there anything hands-on, or is it all theory?It is guidance, not tooling — there is no runnable agent platform here. Hands-on cloud experiences around sandbox, runtime, AI gateway, state storage, observability, and evaluation are explicitly listed as planned future work, not something you can try today.
It is guidance, not tooling — there is no runnable agent platform here. Hands-on cloud experiences around sandbox, runtime, AI gateway, state storage, observability, and evaluation are explicitly listed as planned future work, not something you can try today.Who is it written for — just engineers?Six personas, each with suggested chapters and takeaways: agent/AI app developers (build, run, tuning), architects and platform engineers, tech leads and R&D managers, product and business owners, security/quality/ops staff, and researchers.
Six personas, each with suggested chapters and takeaways: agent/AI app developers (build, run, tuning), architects and platform engineers, tech leads and R&D managers, product and business owners, security/quality/ops staff, and researchers.Why a new whitepaper instead of updating the 2025 architecture one?After the September 2025 AI-native app architecture whitepaper, the team saw focus shift to three new challenges: turning probabilistic agents into reliably productive tools, deploying them at scale with stable security, performance and cost, and moving them from isolated pilots into core business processes — so the structure was rebuilt around those.
After the September 2025 AI-native app architecture whitepaper, the team saw focus shift to three new challenges: turning probabilistic agents into reliably productive tools, deploying them at scale with stable security, performance and cost, and moving them from isolated pilots into core business processes — so the structure was rebuilt around those.What does 'Harness' mean in this book?It is the layer that runs the model. Chapter 3 compares mainstream ways to build it — high-code frameworks, productized harnesses, Managed Agents, and agent cloud products — and draws the responsibility boundary between model and platform; chapters 4-6 then cover task orchestration, context and state, and controlled actions.
It is the layer that runs the model. Chapter 3 compares mainstream ways to build it — high-code frameworks, productized harnesses, Managed Agents, and agent cloud products — and draws the responsibility boundary between model and platform; chapters 4-6 then cover task orchestration, context and state, and controlled actions.How does it handle agent security and governance?A dedicated governance section (chapters 13-16) covers observability and audit, security against prompt injection and data exfiltration with per-action authorization, lifecycle management of AI assets like prompts, skills and MCP servers, and pre-launch agent simulation testing.
A dedicated governance section (chapters 13-16) covers observability and audit, security against prompt injection and data exfiltration with per-action authorization, lifecycle management of AI assets like prompts, skills and MCP servers, and pre-launch agent simulation testing.Are there real company case studies?Yes — the practice section (chapters 25-29) includes Geely's automotive intelligent ops, Tastien's 10,000-store chain ops loop, PolarDB-X alert-to-auto-fix, MiniMax's long-term memory data platform, Bilibili's content insights, plus award-winning projects from the GOAI Agent Infra competition.
Yes — the practice section (chapters 25-29) includes Geely's automotive intelligent ops, Tastien's 10,000-store chain ops loop, PolarDB-X alert-to-auto-fix, MiniMax's long-term memory data platform, Bilibili's content insights, plus award-winning projects from the GOAI Agent Infra competition.What are similar projects?ai-system-design, ai-agents-the-definitive-guide, vibe-coding-cn, awesome-harness-engineering, claude-code-book, ai-agent-book, claude-code-ultimate-guide, zhijian-ai-bluebook-workbuddy-harness
ai-system-design, ai-agents-the-definitive-guide, vibe-coding-cn, awesome-harness-engineering, claude-code-book, ai-agent-book, claude-code-ultimate-guide, zhijian-ai-bluebook-workbuddy-harnessWhich topics is it in?agent-governance, governance, ai-security, context-engineering, architecture, harness, best-practices, evaluation, enterprise, observability, book, artificial-intelligence
agent-governance, governance, ai-security, context-engineering, architecture, harness, best-practices, evaluation, enterprise, observability, book, artificial-intelligence