Published on 28 October as part of Alibaba Cloud's AI Agent Handbook, the account traces the Cloud Communications core product engineering team's path: code completion and generative IDE tools in 2024; standardised workflows with standards and domain knowledge codified into constitutions and Skill resources in 2025; and an agent-based operating model moving toward end-to-end delivery in 2026. Its central observation is that coding is one step in delivery, and the manual steps before and after it cap overall throughput, so AI has to own the whole path once a requirement is filed. On multi-agent handoff the team reports a missing-context problem - the design agent documents architecture and middleware while the testing agent needs system boundaries - and answers it by treating collaboration as a production line, with standards and explicit inputs and outputs at every stage. On permissions it chose the digital-twin model over the digital-employee model: an agent runs under the permissions of the person it represents, so its commits are attributed to that person and problems remain traceable.
Our takeThe digital-twin choice is the technically interesting one, because it turns accountability into an access-control problem the organisation already knows how to solve rather than inventing a new class of actor. What the piece does not give is failure data: how often a stage emits a contract-compliant but wrong artefact, and what the review step actually catches. Pipeline contracts fix handoff semantics, not correctness.