Alibaba Cloud Releases AI Agent Handbook and 2026 Agent Developer Survey Report
English text is being prepared.
Alibaba Cloud's open-source "AI Agent Handbook" includes a "2026 Agent Developer Survey Report," showing that about 46% of teams have completed or are developing Agents, only about 18% have actually deployed to production, and fewer than 10% iterate continuously. The report points out that the bottlenecks are mainly in engineering support: about 90% of enterprises have clear needs for context and memory, multi-model routing and automatic fallback are listed by about 63% as the most needed gateway capabilities, and for evaluation, about 55% use manual spot checks while about 6%–8% use automatic evaluation based on execution traces. In terms of architecture, single-Agent accounts for about 40%, multi-Agent about 42%, and Human-in-the-Loop as the primary architecture about 15%; Coding is the first scenario to scale.
Why read this
AI score 60
The survey data shifts the Agent bottleneck from model capability to engineering support such as context, routing, and evaluation, which teams can compare against their own stage.
Topics
View original postAlibaba Cloud Open-Sources "AI Agent Handbook": Using Harness to Push Enterprise Agents from Pilots to Governable Production A full-lifecycle practice handbook for enterprise-grade Agents: architecture, building, running, governance, and tuning, plus surveys and case studies. Open-source address: https://t.co/CzC395KiJq Why was the handbook rewritten? The preface summarizes three new challenges: turning probabilistic intelligence into reliable productivity that can handle critical tasks; building infrastructure that can be deployed at scale in terms of stability, security, performance, and cost; and getting Agents into core business processes rather than stopping at isolated pilots. The writing goals also include unifying terminology (Harness, Runtime, Agent, Workflow, etc.) to reduce the confusion caused by "the same word meaning different things to different people" being amplified by models. 2026 survey results · Investment is a consensus: about 46% have completed development or are in development (about 36% last year); about 18% have actually deployed to production; continuous iteration is even lower. Forms coexist: single-Agent about 40%, multi-Agent about 42%, and HITL as the primary architecture about 15%—autonomy is a tunable parameter, not a replacement of old by new. Coding scales first: about 57% procure Coding tools, about 78% use general Agent frameworks, and nearly 40% use both; the market is fragmented, and portable engineering abstractions are more stable than betting on a single vendor. Difficulties after building: context/memory needs about 90%; multi-model routing and cost governance about 63%; for evaluation, manual spot checks about 55%, and automatic evaluation based on traces less than about 8%. Success rate: about one in ten reach 90% or above; about 16% have never quantified their success rate at all. Teams with an evaluation system have a noticeably higher proportion of high success rates than those without (compared in the text). MCP: interest or implementation totals about 37%, and private deployment within enterprises is only about 9%; what is missing is a registry, identity authentication, version canary releases, auditing, and a gateway, not "whether people understand the protocol." Maturity L1–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/L4. Overall structure of the handbook · Architecture (1–2): Agentic Application boundaries, enterprise maturity, reference architecture (components / platform responsibilities / lifecycle) · Building (3–6): with Harness as the core: paradigms and responsibility boundaries; task orchestration and long-horizon progression; context, state, Memory/Knowledge/Skill; controlled actions (Function Calling, MCP, A2A, HITL, verification loop) · Running (7–12): Runtime and sandboxing; state storage and semantic assets; AI gateway unifying LLM/MCP/Agent traffic; asynchrony and automation; Multi-Agent; distributed communication · Governance (13–16): observability and cost attribution; security (injection, authentication, high-risk authorization, egress); Prompt/Skill/MCP/Agent asset management; pre-launch simulation and quality verification · Tuning (17–24): model side (SFT, Agentic RL, distillation, etc.) and Agent-side data flywheel: Trace→Trajectory, golden datasets, Badcase, controlled self-evolution, edge runtime · Practice (25–29 + special topics): R&D efficiency, design engineering, operations and enterprise IT, customer sales operations, GOAI Agent Infra; plus a special topic on the Jev decision model and a special topic on Personal Agent / Agent as a Service · Outlook (30): from Agentic Application to Agentic OS
View original
阿里云开源《AI Agent Handbook》：用 Harness 把企业 Agent 从试点推到可治理生产 企业级 Agent 全生命周期实践手册：架构、构建、运行、治理、调优，外加调研与案例，开源地址： https://t.co/CzC395KiJq 手册为何重写？ 前言归纳三大新挑战：让概率智能变成能扛关键任务的可靠生产力；在稳定、安全、性能、成本上做成可大规模部署的基础设施；让 Agent 进入核心业务流程，而不是停在孤岛试点。写作目标还包括统一术语（Harness、Runtime、Agent、Workflow 等），减少「同一词各说各话」被模型放大后的混乱。 2026 年调研结果 · 投入成共识：已开发完成或开发中合计约 46%（去年约 36%）；真正部署到生产约 18%；持续迭代更低。 形态共存：单 Agent 约 40%、多 Agent 约 42%、以 HITL 为主要架构约 15%——自主度是可调参数，不是新旧替代。 编码最先规模化：采购 Coding 工具约 57%，通用 Agent 框架约 78%，近四成两者并用；市场碎片化，可迁移工程抽象比押单一厂商更稳。 构建后难点：上下文/记忆需求约 90%；多模型路由与成本治理约 63%；评估上人工抽查约 55%，基于轨迹自动评估不足约 8%。 成功率：达 90% 以上约一成；约 16% 根本没量化过成功率。有评估体系的团队，高成功率比例明显高于无评估者（文中对比）。 MCP：关注或落地合计约 37%，企业内私有部署仅约 9%；缺的是注册中心、身份鉴权、版本灰度、审计与网关，不是「懂不懂协议」。 成熟度 L1–L4：辅助生成 → 受控自动化 → Agentic Execution → Managed & Optimizing；全书后半部基本在回答如何从 L2 往 L3/L4 走。 手册整体结构 · 架构（1–2）：Agentic Application 边界、企业成熟度、参考架构（组件 / 平台责任 / 生命周期） · 构建（3–6）：以 Harness 为核心：范式与责任边界；任务编排与长程推进；上下文、状态、Memory/Knowledge/Skill；受控行动（Function Calling、MCP、A2A、HITL、验证闭环） · 运行（7–12）：Runtime 与沙箱；状态存储与语义资产；AI 网关统一 LLM/MCP/Agent 流量；异步与自动化；Multi-Agent；分布式通信 · 治理（13–16）：可观测与成本归因；安全（注入、鉴权、高危授权、出域）；Prompt/Skill/MCP/Agent 资产管理；上线前仿真与质量验证 · 调优（17–24）：模型侧（SFT、Agentic RL、蒸馏等）与 Agent 侧数据飞轮：Trace→Trajectory、黄金数据集、Badcase、受控自进化、边缘运行时 · 实践（25–29 + 专题）：研发效能、设计工程、运维与企业 IT、客户销售运营、GOAI Agent Infra；另有 Jev 决策模型专题、Personal Agent / Agent as a Service 专题 · 展望（30）：从 Agentic Application 走向 Agentic OS
Body
阿里云开源的「AI Agent Handbook」有一份「2026 Agent 开发者调研报告」
做 Agent 已成共识，近半团队在做，真正进生产仍是少数（~18%）；卡点主要是工程配套（上下文与记忆、多模型路由与成本、评估与可观测），不是「要不要做」。单 Agent / 多 Agent / 人在回路会长期并存；Coding 最先规模化，但工具与框架高度碎片化，可迁移的抽象比押注单一厂商更值钱。 https://t.co/oDvCxODDi8
投入升温，生产化仍窄 「暂无计划」从去年 22% 降到 15%；「调研/计划」从 47% 降到 25%；「已完成或开发中」从 36% 升到 46%；真正部署到生产约 18%。持续迭代的不到 10%。上线差距往往大于开发意愿差距：上海大型企业已上线约 45%、小微约 16%；深圳约 38% 对 17%。报告判断：差的不是愿不愿意投，而是能否持续演进 Agent 的组织协作、治理、优化与沉淀。并挂上白皮书 L1–L4：多数仍在辅助生成或受控自动化，Agentic Execution 只在部分场景成立，Managed & Optimizing 仍少。
三种形态共存，不是新旧替代 主要架构：单 Agent 约 40%，多 Agent 约 42%，另有约 15% 把 Human-in-the-Loop（关键步骤必须人确认）当作生产架构原则，而不只是外挂审批。自主度是可调参数；Chat/RAG、Workflow、Copilot、Agent 与 Managed Agentic Application 常在同一家公司并存，取决于任务确定性与容错空间。
多 Agent 痛点：状态与上下文在协作中衰减约 60%；死锁雪崩、成本失控也被点名。最希望补齐：研发—测试闭环约 65%，端侧与开源运行底座约 73%，验证成本已成为负担，运行底座也要自主可控。
Coding 先跑通，供应商却很碎片 采购 Claude Code、Codex、Cursor、Qoder 等 Coding 工具约 57%；用通用 Agent 框架约 78%；近四成两边都用。Coding 被认为是目前唯一规模化付费场景：反馈清晰（编译/测试）、边界清楚（仓库与工作区）、错误可回滚。前十供应商呈平缓下降而非陡峭长尾，多框架并行试用仍常见；围着单一框架定企业内部标准风险偏高。办公 Agent 兴起后，编程 Agent 与通用框架边界也在糊。
落地场景排序：员工效能、代码工程、数据分析更广；引入企业核心业务流程不到 40%，优先容错大、能人工兜底的场景，严肃流程要更完整的治理。
构建后三大难点：信息、治理、行动 约 90% 企业对上下文与记忆有明确需求。「无明显痛点」不到 10%。记忆侧：检索不准约 54%，更新与遗忘机制缺失约 48%，窗口大小只排第三：问题在写入、淘汰与检索策略，不在窗口够不够大。工具侧：工具过多难选约 52%，幻觉调用约 50%，组织和描述质量比堆数量更影响成功率。 MCP：已关注或落地合计约 37%，企业内私有部署仅约 9%。真正跑起来需要私有注册中心、统一身份鉴权、版本灰度、审计，以及与网关集成，这些协议本身不提供。约 28% 明确要 MCP 服务管理与注册中心，和已落地比例接近，说明需求正从认知转向工程。
网关与可观测：路由第一，成本归因很早 多模型路由与自动降级被约 63% 列为最需要的网关能力（单项最高）：按任务选模型、主模型挂了降级、按成本与延迟动态选。这类事很难在各应用里各自硬编码，需要统一流量入口。可观测诉求里，成本归因约 50%，仅次于全链路追踪，高于审计合规与语义质量监控，Token 还不大时就开始做可见性与分摊，偏预防性建设。
评估：多数还在人工抽查 评估方法：人工抽查约 55%；基于运行轨迹自动评估约 6%–8%；公开 Benchmark 不到 5%；尚无系统性评估约 28%。LLM-as-Judge、A/B、离线回归三类里用上任一种的约 34%。
任务成功率：≥90% 的企业约一成；≥70% 约 55%；约 16% 从未量化过成功率，等于说不清 Agent 是否在正常工作。有评估手段的企业，成功率 ≥70% 大约是无体系企业的两倍；已上线并持续迭代的企业里这一比例约 84%。报告串成闭环：评估支撑迭代，迭代抬可靠性，可靠性又让评估值得持续做；缺评估则难定位、难证明改动有效，容易长期停在试点，与「近半在开发、约五分之一上线」相互印证。
总结一下 投入侧已过可行性验证；卡点不是意愿也不是单纯模型能力，是 Harness 层可开箱的工程配套。形态与工具都碎片化，Hybrid 是常态。白皮书目录按调研痛点排布：架构定形态与边界，构建用 Harness 作不绑框架的组合，运行回应沙箱/网关/异步/多 Agent 通信，治理把质量从人工兜底做成可规模化机制，调优沿 Trace→Trajectory 让系统持续变好。