AI Agent HandBook: Alibaba Cloud Enterprise Agent White Paper

Alibaba Cloud's Open-Source Playbook for Enterprise Agents

AI Agent HandBook (github.com/aliyun/ai-agent-handbook, Apache-2.0, created Sep 11, 2026) is a white paper from Alibaba Cloud distilling enterprise agent deployment experience along the agent lifecycle — architecture, building, runtime, governance, and optimization — currently at 516 stars. It is not another framework or SDK but an open, continuously evolving engineering handbook: 30 chapters, a 2026 agent developer survey, and enterprise case studies, all maintained as Markdown in the repo and open to community collaboration.

It builds on the AI-Native Application Architecture White Paper Alibaba Cloud published in September 2025. With model capability no longer the bottleneck, the real challenges have become engineering (turning probabilistic intelligence into reliable productivity so agents can take on critical tasks) and scaling (stability, security, performance and cost as agents move from isolated experiments to deployable intelligent infrastructure). The book is rewritten around these two themes.

Structure: Six Parts, 30 Chapters

Part	Chapters	Topics
Architecture	1–2	Agentic Application definition and boundaries, enterprise maturity, reference architecture (component / platform-responsibility / lifecycle views)
Building	3–6	Harness construction patterns and responsibility boundaries; task orchestration and long-horizon collaboration; context, state and reusable capability assets; controlled execution and verification
Runtime	7–12	Runtimes and sandboxes; state storage and semantic assets; AI gateways and unified traffic governance; async agent tasks; multi-agent orchestration; distributed agent communication
Governance	13–16	Observability; security (prompt injection, identity, per-action validation, high-risk authorization, data egress); discovery and management of prompts/skills/MCP/agents; behavior generation and pre-release validation
Optimization	17–24	A continuous-improvement data flywheel of Trace / Trajectory / golden datasets / evaluation experiments
Practice & Outlook	25–30	Enterprise cases (ABACI kernel-patch testing, Kitta domain-specific code review, Bilibili cross-platform content insights); from Agentic Application to Agentic OS

Why It Matters: The Harness Perspective

For readers in the harness/agent-infrastructure space, the core claim lives in chapters 3–5: understand the responsibility boundary between the Model and the Harness — don't attribute every failure to model capability. The guide explicitly covers Agent loops, task state machines, planning and stage gates, delegation, asynchronous continuation and completion evidence (ch. 4); Context builders, compression and offloading, Session/Task state, workspaces, memory, knowledge and skills (ch. 5); Action planes, Function Calling, MCP, A2A, environment contracts, permissions and HITL (ch. 6). Chapters 7–9 then assemble a production execution view spanning runtimes and sandboxes, Event Log / Checkpoint / workspace snapshots / artifacts, and AI-gateway governance (identity, permissions, budgets, routing, audit) across LLM, MCP and agent traffic.

The role-based reading paths are practical too: developers read Building/Runtime/Optimization for harness, context, state, tools, sandboxes, trajectories and the evaluation loop; architects read Architecture/Runtime/Governance for extensible agent infrastructure design; tech leads read Architecture/Governance/Practice to judge application form, maturity, investment boundaries and production risk; security/QA/ops readers get observability, audit, release validation and incident attribution.

Notable Points

• Open evolution, not a frozen document: the roadmap commits to revisiting conclusions as models, harnesses, protocols, runtimes, multi-agent systems and Agentic OS evolve, with community collaboration on guidelines, case templates, terminology and review processes.
• Includes a 2026 agent developer survey covering enterprise agent development, productionization, architecture choices, tooling, governance and evaluation — readable on its own.
• Real business cases: software engineering (kernel patch testing, domain code review) and customer operations (Bilibili content insights), not toy demos.
• Language note: chapter content is currently in Chinese; the English README is a guide, not a full translation.

Relevance to RobotWorld

The site already tracks harness-flavored projects like langchain-ai/deepagents, XiaomiMiMo/MiMo-Code and Untrivial-ai/agent-orchestrator — those show how to build a harness; this handbook answers how enterprises run, govern and tune one once it exists. The two perspectives complement each other nicely.

Resources

• Repo: https://github.com/aliyun/ai-agent-handbook (Apache-2.0)
• Predecessor: AI-Native Application Architecture White Paper (developer.aliyun.com/ebook/8479)
• Keywords: Agentic Application, Agent Harness, context engineering, MCP, A2A, agent governance, evaluation loop, Agentic OS