Alibaba Cloud Releases AI Agent Handbook and 2026 Agent Developer Survey Report

Alibaba Cloud's open-source "AI Agent Handbook" includes a "2026 Agent Developer Survey Report"

Building Agents is now consensus; nearly half of teams are doing it, but only a minority (~18%) have actually reached production; the blockers are mainly engineering support (context and memory, multi-model routing and cost, evaluation and observability), not "whether to do it." Single Agent / Multi-Agent / Human-in-the-Loop will coexist long-term; Coding is the first to scale, but tools and frameworks are highly fragmented, and transferable abstractions are more valuable than betting on a single vendor. https://t.co/oDvCxODDi8

Investment is heating up, but productionization remains narrow "No plans yet" fell from 22% last year to 15%; "researching/planning" fell from 47% to 25%; "completed or in development" rose from 36% to 46%; about 18% are truly deployed to production. Fewer than 10% are continuously iterating. The deployment gap is often larger than the willingness-to-develop gap: large enterprises in Shanghai have about 45% live, versus about 16% for small and micro enterprises; Shenzhen is about 38% versus 17%. The report concludes: what's lacking is not willingness to invest, but the organizational collaboration, governance, optimization and accumulation needed to continuously evolve Agents. It also attaches the whitepaper's L1–L4: most are still at assisted generation or controlled automation, Agentic Execution holds only in some scenarios, and Managed & Optimizing remains rare.

Three forms coexist, not a replacement of old by new Main architectures: single Agent about 40%, multi-Agent about 42%, and another about 15% treat Human-in-the-Loop (key steps must be confirmed by a human) as a production architecture principle, not just an external approval layer. Autonomy is a tunable parameter; Chat/RAG, Workflow, Copilot, Agent and Managed Agentic Application often coexist within the same company, depending on task determinism and tolerance for error.

Multi-Agent pain points: state and context decay by about 60% during collaboration; deadlock avalanches and runaway costs are also called out. Most wanted improvements: the development–testing loop at about 65%, on-device and open-source runtime foundations at about 73%; verification cost has become a burden, and the runtime foundation also needs to be autonomous and controllable.

Coding runs first, but vendors are fragmented About 57% procure Coding tools such as Claude Code, Codex, Cursor, Qoder; about 78% use general Agent frameworks; nearly 40% use both. Coding is considered the only scaled paid scenario so far: clear feedback (compilation/tests), clear boundaries (repository and workspace), and errors can be rolled back. The top ten vendors show a gentle decline rather than a steep long tail, and parallel trials of multiple frameworks remain common; setting enterprise internal standards around a single framework carries relatively high risk. As office Agents rise, the boundary between coding Agents and general frameworks is also blurring.

Deployment scenario ranking: employee productivity, code engineering, and data analysis are broader; adoption into core enterprise business processes is under 40%, prioritizing scenarios with high error tolerance and human fallback, while serious processes need more complete governance.

Three major difficulties after building: information, governance, action About 90% of enterprises have clear needs for context and memory. "No obvious pain points" is under 10%. On the memory side: inaccurate retrieval about 54%, missing update and forgetting mechanisms about 48%, and window size ranks only third: the problem lies in writing, eviction and retrieval strategies, not whether the window is large enough. On the tool side: too many tools to choose from about 52%, hallucinated calls about 50%; organization and description quality affect success rates more than piling on quantity. MCP: combined attention or implementation about 37%, private deployment within enterprises only about 9%. Getting it truly running requires a private registry, unified identity authentication, version canary releases, auditing, and integration with gateways, none of which the protocol itself provides. About 28% explicitly want MCP service management and a registry, close to the implemented share, indicating demand is shifting from awareness to engineering.

Gateway and observability: routing first, cost attribution very early Multi-model routing and automatic fallback are cited by about 63% as the most needed gateway capability (the highest single item): selecting models by task, falling back when the primary model fails, and dynamically choosing by cost and latency. This is hard to hardcode separately in each application and requires a unified traffic entry point. Among observability demands, cost attribution is about 50%, second only to full-link tracing and higher than audit compliance and semantic quality monitoring; visibility and cost allocation are being built while Token usage is still small, a preventive approach.

Evaluation: most still rely on manual spot checks Evaluation methods: manual spot checks about 55%; automatic evaluation based on execution traces about 6%–8%; public Benchmarks under 5%; no systematic evaluation about 28%. Among LLM-as-Judge, A/B, and offline regression, about 34% use at least one.

Task success rate: about one in ten enterprises at ≥90%; about 55% at ≥70%; about 16% have never quantified success rate, meaning they cannot say whether the Agent is working properly. Among enterprises with evaluation methods, the share at ≥70% success rate is about twice that of enterprises without a system; among enterprises already live and continuously iterating, this proportion is about 84%. The report forms a closed loop: evaluation supports iteration, iteration raises reliability, and reliability makes evaluation worth continuing; without evaluation it is hard to locate problems and prove changes are effective, and easy to stay in pilot for a long time, which corroborates "nearly half in development, about one-fifth live."

To summarize On the investment side, feasibility has been validated; the blocker is neither willingness nor pure model capability, but out-of-the-box engineering support at the Harness layer. Forms and tools are both fragmented, and Hybrid is the norm. The whitepaper's table of contents is arranged around survey pain points: architecture defines form and boundaries, building uses Harness as a framework-agnostic composition, running addresses sandbox/gateway/async/multi-Agent communication, governance turns quality from manual fallback into a scalable mechanism, and tuning follows Trace→Trajectory to make the system continuously better.