Launching a Multi‑Agent AI Platform in One Week with Alibaba Cloud AgentTeams and AI Gateway

XinYongZhongHe, one of China’s earliest accounting firms, has been digitizing and integrating AI into professional services, knowledge acquisition, business processing, and decision making.

Enterprise AI needs now extend beyond single Q&A bots to address multi‑agent collaboration, unified model governance, system integration, security, and cost control.

To meet these needs, XinYongZhongHe collaborated with Alibaba Cloud, adopting AgentTeams for multi‑agent governance and the AI Gateway for model access, jointly building an enterprise‑grade AI platform.

Multi‑Agent Collaboration

Using AgentTeams, the firm created specialized agents for different business scenarios and a management agent that identifies task types, dispatches to appropriate agents, and aggregates results, enabling complex multi‑step tasks.

Agents built include:

Policy & Process Q&A Agent – connects to internal knowledge bases to answer questions about regulations and procedures.

OA Integration Agent – interfaces with internal office systems for time‑sheet queries and submissions.

Enterprise Information Query Agent – provides data on company basics, shareholders, relationships, and risks.

Document Processing & Reporting Assistant Agent – helps summarize documents, extract material, and generate report frameworks.

Unified Model and AI Asset Management

AgentTeams offers centralized management of models, MCP services, Skills, and agent templates, allowing internal systems to be exposed as standard capabilities for agents, enabling not only question answering but also authorized information queries, system calls, and task execution.

Fine‑grained statistics track token usage and call volume per department, agent, and model, supporting resource allocation, usage analysis, and budget management.

AI Gateway: Centralized Model Invocation and Security

The AI Gateway serves as a unified entry for model calls, providing monitoring, authentication, traffic control, and cost analysis.

Multi‑model scheduling ensures service stability: when a primary model experiences latency or errors, predefined policies switch to a backup model.

Security mechanisms such as sensitive‑information detection, content safety checks, and access control reduce data‑risk during model invocation.

Observability and Enterprise AI Governance

The platform records agent usage, model calls, task execution, and anomalies, generating audit logs for troubleshooting, security management, and internal governance.

Unified identity authentication, permission control, credential management, and resource isolation ensure agents operate within authorized scopes.

Implementation Outcome

In the final stage, both teams completed system integration, environment configuration, application onboarding, and testing within one week, launching on July 6. The platform now offers a unified AI service entry for employees, allowing conversational access to knowledge, information queries, and business capabilities, lowering the barrier compared to traditional menus.

The unified agent management and AI Gateway establish a technical foundation for future agents and scenarios, moving AI from answering questions to executing work.

Future Directions

XinYongZhongHe plans to expand agents for policy services, data analysis, cross‑system task coordination, project‑management assistance, and professional business support, further integrating AI into internal workflows.