Selecting and Running an Agent Data Flywheel Platform: The Zeus Practice at 1688
Zeus is 1688's experiment in an AI-driven organizational management brain. This article documents how the team wired Alibaba Cloud's AgentLoop into it, moving from "did the call succeed" to observation, evaluation and regression testing.
Zeus connects business knowledge, data and multi-role agents so AI can analyze against business goals, coordinate execution and follow up continuously. Before adoption, the team's complaint was "examining an elephant by touch": a successful API call says nothing about whether the task actually completed, where the deviation occurred, or whether a fix will hold. AgentLoop adds three layers — observability (see how one request actually executes), evaluation (score against task completion, not call success), and regression validation (failed samples become inputs to the next test run, organized as test sets, regression sets and baseline sets). Implementation uses ARMS GenAI Java Agent probes, datasets of inputs plus outputs plus process traces, and selectable evaluators scored by dimension. Two cases are replayed in detail: a document read that never refreshed, and a subtask result that never came back. The recommendation is to start from the smallest closed loop and grow outward.
The full text of this article is currently available in Chinese only. Read the original