Building a Scalable 800-Agent Digital Workforce with Alibaba Cloud AgentCore in 39 Days

Wanlian Zhice, an intelligent enterprise opportunity mapping platform under Wanlian Yida Group, partnered with Alibaba Cloud to upgrade its AI assistant "Xiaotu" from a simple Q&A bot into a full-fledged multi-agent digital employee team. The new system, dubbed the "Business Opportunity Think Tank," comprises one Leader agent, four expert Worker agents, and a pool of approximately 800 configurable agents that run on-demand. Built on Alibaba Cloud's AgentCore for multi-agent collaboration and AgentLoop for observation and evaluation, the solution integrates directly into Wanlian Zhice's own mobile application without migrating existing systems.

Business Context and Challenges

Wanlian Zhice provides enterprise decision-making tools based on geographic, enterprise-profile, and supply-chain data. Its AI entry point, "Xiaotu," previously handled only conversational Q&A. The product team wanted digital employees to execute end-to-end business tasks — information gathering, analysis, and report generation — around concrete business goals. Four key challenges emerged:

Business division of labor: Supplier information collection, risk analysis, and result compilation are interdependent and require clear role definitions and upstream/downstream relationships.

Application integration: Multi-agent capabilities had to be embedded into the existing app without system migration, aligning tenants, teams, sessions, and tasks.

Long-running task experience: Users need real-time progress visibility, confirmation checkpoints, and the ability to leave and resume later.

Cost and quality control: Assigning a dedicated agent team per user would incur idle costs if instances were always resident, and single-turn answers cannot reflect overall task quality.

Solution Architecture: AgentCore + AgentLoop

AgentCore: Multi-Agent Collaboration Runtime

AgentCore adopts a Leader-Worker organization pattern. The Leader plans tasks, dispatches work, and aggregates results; Workers encapsulate business skills and tools. For example, in a supplier research scenario:

Xiaotu (Leader) asks clarifying questions: region, preferences, quantity.

Leader dispatches a Supply Chain Worker to find suppliers and a Risk Control Worker to assess risks.

Leader aggregates the findings into a structured report.

Tasks run primarily in serial and scheduled modes. Model selection is role-specific: the Leader uses a planning-oriented model, while expert Workers use execution-oriented models. The generated analysis assists human decision-makers; final actions such as purchase orders and contract signing remain manual.

Figure 1: Role division in supplier research

AgentLoop: Task-Level Observation and Evaluation

AgentLoop unifies model calls, tool executions, and run trajectories, aggregating them by task ID. This enables full visibility into the Leader's reasoning and each Worker's actions. Evaluation follows a rigorous loop:

Fixed benchmark datasets for version regression testing.

Online experiments with isolated samples.

Automatic removal of error and retry samples before quality judgment.

Continuous Observe → Evaluate → Optimize cycle.

Integration into Wanlian Zhice's Application

The platform calls cloud-hosted agents via API. Users enter a Team, initiate tasks within a session, and Workers invoke existing business capabilities through Skill and MCP (Model Context Protocol). General capabilities like web search use Alibaba Cloud's Bailian MCP, while proprietary industrial-chain, enterprise, and public-opinion data are served by internal services. The mobile app accesses the system through a cloud-native API gateway that handles unified authentication and maps execution roles to digital employee identities.

Figure 2: Business access model from tenant to capabilities

Long-Task Experience and Operational Stability

Streaming reasoning: The Leader's thought process is streamed in real-time; missing information triggers proactive follow-up questions.

Cloud persistence: Tasks run in the cloud; users can review progress and artifacts after leaving the session. Scheduled tasks are already live.

On-demand resources: Minimum instance count is zero; a warm-up pool reduces first-response latency. Capacity planning and search-call optimization ensure peak stability.

Data separation: External display data and internal runtime data are handled separately. User identifiers are anonymized; content, reasoning traces, and tool calls are processed independently.

Business Outcomes

800 Agents Running On-Demand

Production configuration includes roughly 800 agents that start and stop based on actual sessions and tasks, with a minimum instance count of zero. This figure represents the configured business scale, not concurrently resident compute instances.

Personalized "Business Opportunity Think Tank" per User

After business process analysis, Wanlian Zhice defined a comprehensive opportunity evaluation workflow: for a single bidding notice, the team analyzes from sales, procurement, risk, and strategy perspectives. In supplier risk scenarios, the system not only detects risks but also recommends alternative suppliers. The deployed Think Tank assigns each user a dedicated Team:

Leader — Xiaotu: Understands requirements, asks clarifying questions, decomposes tasks, aggregates results.

Sales Consultant: Finds customers (market insight & opportunity acquisition).

Supply Chain Director: Finds suppliers (supply chain organization & fulfillment).

Risk Expert: Assesses risks (risk identification & compliance).

Strategy Expert: Conducts research (strategic judgment & decision support).

The original Xiaotu capabilities were refactored into the skills and tools of these four experts. Xiaotu now orchestrates them into a complete business collaboration flow.

Figure 3: Logical relationship between business app, agent runtime, and supporting capabilities

39 Days from Concept to Production

From the first Agent business deployment on AgentCore to the supplier research workflow entering actual operation, Wanlian Zhice completed the upgrade in 39 days, significantly accelerating business innovation velocity.

Verifiable Task Processes

Trajectories of the Leader and all Workers are linked by task ID, allowing inspection of role division, skill invocations, and report artifacts against expectations — not just the final answer.

Usage-Based Cost Accounting

Costs cover platform instances, agent compute, model tokens, AI gateway, and storage. Accounting combines task duration, input/output volumes, and cache usage, enabling identification of high-consumption patterns such as long contexts and repeated calls.

Future Directions

The team will continue improving task reliability, task-level evaluation, first-response latency, and resource efficiency. They are also building B-end management capabilities for Wanlian Zhice, using runtime data — rather than agent counts or feature checklists — to measure progress.