46% of Teams Are Building Agents, 18% Dare Let Them Touch Production

Alibaba Cloud has open-sourced an "AI Agent Handbook," 30 chapters, with a GitHub repository under the Apache 2.0 license and also an online reading site. The content centers on the entire lifecycle of Agents—architecture, building, running, governance, and tuning—plus a 2026 developer survey and a set of enterprise cases.

Repository address: https://github.com/aliyun/ai-agent-handbook

The preface explains why it was rewritten. In September last year they published the "AI-Native Application Architecture White Paper," which was about how to quickly build agents. A year later, the market's focus has shifted, and it comes down to three sentences: make probabilistic intelligence into reliable productivity that can handle critical tasks; turn stability, security, performance, and cost into infrastructure that can be deployed at scale; and let Agents enter core business processes instead of stopping at isolated pilots. Incidentally, they also unified a set of terms—what Harness, Runtime, Agent, and Workflow each refer to—aligning them first, because when the same word means different things to different people, the model amplifies the confusion.

There are several sets of survey data worth calling out separately.

Completed or in-development together account for about 46%; last year this number was 36%. Those actually deployed to production are about 18%. The proportion continuously iterating is even lower. In terms of form, single Agent is about 40%, multi-Agent about 42%, and HITL as the primary architecture about 15%. Autonomy is a tunable parameter, unrelated to new versus old. Coding is the first scenario to scale; procurement of general Agent frameworks is about 78%, Coding tools about 57%, and nearly 40% use both.

After building, the difficulties concentrate in two areas: demand for context and memory is about 90%; multi-model routing and cost governance about 63%. On evaluation, manual spot checks are about 55%, while trajectory-based automated evaluation is less than 8%. About one in ten have success rates above 90%, and about 16% have not quantified the success rate at all. MCP interest or implementation totals about 37%, with private deployment inside enterprises accounting for only 9%.

The gap between 46% and 18% drew the most discussion in the comments. One commenter said they stared at that line because it matched reality: the demo ends to full applause, but once actually connected to business it all relies on manual babysitting as a fallback. Later they were forced into a crude rule: if it wasn't clearly written whether it passed or not, it was not allowed into core processes. Clumsy, but more useful than building another test suite.

The problem with the evaluation numbers is more hidden. 90% of teams say they need context and memory, yet fewer than 8% have trajectory-based automated evaluation. Many teams are changing prompts, but cannot answer how much the success rate changed with this change. Once multi-Agent reaches 42%, it gets even harder: manual spot checks only see the final output, and cannot tell which handoff step lost context. A commenter followed up on this: when the governance chapter discusses cost attribution, does it break down token spend by trace down to the level of an individual Agent and a single tool call? That's a good question, and exactly where enterprises really need to do the accounting.

The handbook makes governance a separate section, chapters 13–16: observability and cost attribution; security (injection, authentication/authorization, high-risk authorization, egress); Prompt/Skill/MCP/Agent asset management; pre-launch simulation and quality validation. In community feedback, this section gets the highest marks. One commenter's summary: what makes enterprise Agents hard is never the demo; it's permissions, auditing, and who takes the blame when something goes wrong. Another commenter's approach is to first build a private gateway on entry, bind tool calls tightly to approval workflows and identity authentication, test rework rate with fixed samples before launch, and if audit logs cannot be produced, risk control won't sign off. An even plainer three-point version: write clearly what each AI can touch, have someone accept it when done, and stop and ask a human when encountering an irreversible operation.

There were also doubts. One commenter directly asked whether Alibaba Cloud has any successful FDE cases; if not, what is the point of reading the docs. The handbook does not directly answer this. Another commenter who went through the repo mentioned that among the 18 cases in the practice section, 9 are R&D efficiency; the rest include operations implementations such as Geely Auto and Tastien, as well as design engineering and customer operations.

The handbook's table of contents is roughly as follows: Architecture (Chapters 1–2) defines the boundaries of Agentic Applications and enterprise maturity; Building (Chapters 3–6) centers on Harness and covers task orchestration, long-horizon progression, context and state, Memory/Knowledge/Skill, and controlled actions; Running (Chapters 7–12) covers Runtime and sandboxing, state storage, an AI gateway that unifies three types of traffic—LLM/MCP/Agent, asynchronous tasks, Multi-Agent, and distributed communication; Governance (Chapters 13–16); Tuning (Chapters 17–24) has two tracks: model side and Agent side. The model side is SFT, Agentic RL, and distillation; the Agent side is the data flywheel from Trace to Trajectory, golden datasets, Badcase regression, and controlled self-evolution; Practice (Chapters 25–29 plus two special topics); Outlook (Chapter 30) from Agentic Application to Agentic OS.

Maturity is divided into L1 through L4: assisted generation, controlled automation, Agentic Execution, Managed & Optimizing. The second half of the book is basically answering how to move from L2 to L3 and L4.

The handbook is not short—30 chapters. It is suitable for browsing as a reference book, not for reading from cover to cover. The truly informative judgments are hidden in two places: one sentence is that autonomy is a tunable parameter; the other is that set of survey numbers. Investment is rising, evaluation hasn't kept up, and governance is not yet institutionalized. If these three things don't advance at the same time, the gap between 46% and 18% will not close by itself.