Alibaba Cloud Open-Sources Enterprise Agent Whitepaper: 7 Parts, 30 Chapters on Agent Engineering and Scaling

Alibaba Cloud has open-sourced the 'Enterprise-Grade Agent Whitepaper', an upgraded sequel to the September 2025 'AI-Native Application Architecture Whitepaper', organized by the full lifecycle of 'architecture → build → run → govern → tune', with 7 parts and 30 chapters in total. The whitepaper states its writing motivation is that the market focus has shifted from 'how to quickly build an Agent' to three new challenges: engineering, scaling, and organization.

This motivation itself is noteworthy: it moves the starting point of discussion from 'is the model strong enough' to 'is the supporting infrastructure complete enough'. The whitepaper was written by dozens of frontline engineers at Alibaba Cloud in division of labor, and includes external enterprise cases such as Geely, Tastien, MiniMax, Bilibili, and ShineWing. It should be noted that the scale of authors and depth of cases are both claims from the publisher, and the source does not elaborate on specific content.

The architecture part (chapters 1–2) establishes a cognitive framework, giving six defining features of Agentic Applications: centered on task outcomes, runtime determines part of the execution path, can act on the environment, maintains state across requests, constrained by deterministic mechanisms, observable and evaluable. The role of these six features is to draw boundaries—distinguishing 'called a large model once' from 'actually running an Agent task'.

Complementing this is a four-level maturity model: L1 assisted generation, L2 controlled automation, L3 Agentic Execution, L4 scaled operations. The overarching principle of the whitepaper is 'minimum sufficient architecture', i.e., choosing the lowest complexity acceptable in cost and risk for the task. Inferring from this principle, maturity is not better the higher it is, but must match the completeness of governance—this is an inference drawn from the whitepaper's judgment that 'which maturity level one is at depends on the completeness of governance', not the original wording in the source.