Building a Data-Centric Multi-Agent Platform for Enterprise AI

alibaba · Tech · 2025 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise AI agent platform

Alibaba describes how enterprises can move from isolated single-task agents to data-centric multi-agent systems that connect private data, tools, prompts, feedback, observability, and real-time RAG updates.

Constraints:

enterprise agents need private system and customer data rather than only public model knowledge
agent behavior depends on prompt and data updates that must be changed and observed after launch
model, tool, and account access must be secured and governed
the article is promotional and gives limited methodology behind its performance claims

Architecture

Alibaba's proposed architecture combines Spring-AI-Alibaba for agent construction and orchestration, Higress as an AI-native gateway for models, tools, MCP servers, and security controls, Otel for end-to-end reasoning and retrieval tracing, Nacos for dynamic prompt updates, and RocketMQ for real-time RAG data synchronization. The components are framed as a data flywheel: feedback and evaluation improve private data quality, which improves the agents.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  customer-data[Customer data] --> apache-rocketmq[Apache RocketMQ]
  apache-rocketmq --> rag-knowledge-base[RAG knowledge base]
  customer-data --> mcp-server[MCP Server]
  mcp-server --> higress-gateway[Higress gateway]
  model-providers[Model providers] --> higress-gateway
  higress-gateway --> spring-ai-alibaba[Spring-AI-Alibaba]
  rag-knowledge-base --> spring-ai-alibaba
  nacos-prompt-configuration[Nacos prompt configuration] --> spring-ai-alibaba
  spring-ai-alibaba --> intelligent-diagnostic-system[Intelligent diagnostic system]
  spring-ai-alibaba --> otel-observation-system[Otel observation system]
  rag-knowledge-base --> otel-observation-system
  toolset[Toolset] --> otel-observation-system
  toolset --> higress-gateway

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Spring-AI-Alibaba`	Agent construction and orchestration framework for consultation and anomaly-handling workflows.	`Higress gateway`, `Intelligent diagnostic system`
`Higress gateway`	AI-native API gateway for model access, tool integration, TLS, API-key management, quotas, and data security.	`Spring-AI-Alibaba`, `MCP Server`, `Model providers`, `Customer data`
`MCP Server`	Customer-data access path used by agents through Higress.	`Higress gateway`, `Customer data`
`Otel observation system`	End-to-end tracing and quality analysis for reasoning, search, recall, knowledge-base, RAG, and tool behavior.	`Spring-AI-Alibaba`, `RAG knowledge base`, `Toolset`
`Nacos prompt configuration`	Dynamic real-time prompt and algorithm update mechanism with gray configuration for monitoring changes.	`Spring-AI-Alibaba`
`Apache RocketMQ`	Real-time event and data synchronization layer for keeping RAG data current.	`Customer data`, `RAG knowledge base`
`RAG knowledge base`	Retrieval data layer updated from system and customer data.	`Apache RocketMQ`, `Otel observation system`, `Spring-AI-Alibaba`
`Intelligent diagnostic system`	Alibaba's reported applied solution for consulting issues and anomaly diagnosis.	`Spring-AI-Alibaba`, `Higress gateway`
`Model providers`	Underlying LLMs shielded behind unified protocols, permissions, and disaster-recovery controls.	`Higress gateway`
`Customer data`	Private enterprise data used for inference and continuously updated RAG context.	`Higress gateway`, `MCP Server`, `Apache RocketMQ`
`Toolset`	External tools used by the agent and traced for quality diagnosis.	`Higress gateway`, `Otel observation system`

Data flow

Enterprise data and tools enter through Higress and MCP/search integrations. Spring-AI-Alibaba coordinates agent chat or composer workflows. Otel traces reasoning, search, and recall quality. Nacos pushes prompt updates, while RocketMQ propagates RAG data changes so each inference can use fresher data. Customer feedback and evaluation data feed back into the private data system.

3 supporting quotes

“* Integrate system data and toolsets with one-click through Higress to obtain private customer data. * Complete full chain data quality observation through the Otel observation system.”article, stated

“Access domain data through search tools, and customer data through MCP Server, integrating the complete data needed for inference.”article, stated

“Both system data and customer data are constantly updated. We can sync change events and data in real-time through RocketMQ”article, stated

Why this architecture

The architecture is designed around the claim that LLMs alone are insufficient for enterprise work. Alibaba emphasizes a data-centric multi-agent model where private data access, data-quality observation, prompt updates, real-time RAG refreshes, and security controls are platform capabilities rather than one-off app code.

3 supporting quotes

“Considering generality and specificity, a balance is often needed, and cost effectiveness must also be balanced. The arrival of AGI will still take some time; therefore, we judge that the main future direction will be a data-centric, multi-agent collaboration model.”article, stated

“Building high-quality data and continuously optimizing data quality capabilities will be crucial for the success of these agents.”article, stated

“By continuously optimizing the private high-quality data system on the right through the evaluation data system on the left, we can achieve high-quality alignment between customer demands and data.”article, stated

Models used: DeepSeek · LLMs behind Higress

Scale

Alibaba reports the diagnostic system solves more than 95 percent of consulting issues.
“Using the above technical system, we have established an intelligent diagnostic system through open-source AI experts alongside Alibaba Cloud's native API gateway and Microservices Engine (MSE), solving over 95% of consulting issues and over 85% of anomalies.”article, stated
Alibaba reports the system solves more than 85 percent of anomalies.
“Using the above technical system, we have established an intelligent diagnostic system through open-source AI experts alongside Alibaba Cloud's native API gateway and Microservices Engine (MSE), solving over 95% of consulting issues and over 85% of anomalies.”article, stated
The design targets real-time prompt and RAG data updates rather than static configuration.
“* Dynamically update prompt word data in real-time via Nacos to see optimization effects. * Use Apache RocketMQ to dynamically update RAG data for real-time feedback and optimization.”article, stated

Alternatives considered (roads not taken)

Fixed single-task agents

Rejected because: Alibaba argues the field is shifting toward platforms that support multi-agent collaboration and data-quality optimization.

“efforts have shifted towards building agent platforms and paradigms to enhance multi-agent collaboration, orchestration, and optimization of data quality systems.”article, stated

Waiting for general AGI-style agents

Rejected because: The article says AGI will still take time and positions data-centric multi-agent collaboration as the practical direction.

“The arrival of AGI will still take some time; therefore, we judge that the main future direction will be a data-centric, multi-agent collaboration model.”article, stated

Direct model and tool integration without a gateway

Rejected because: Higress is used to shield underlying model/tool systems and centralize secure data links, account security, quotas, and key management.

“By leveraging Higress to shield underlying models and tool systems, we build secure data links and account security systems.”article, stated

Patterns

`multi-agent-orchestration` (orchestration)

Alibaba frames enterprise AI as a multi-agent architecture that circulates data and tasks rather than isolated single-purpose agents.

“Implement data and task automatic circulation through a multi-agent architecture.”article, stated

`tool-use-orchestration` (orchestration)

Higress exposes models, domain search, customer data, MCP servers, and tools behind a controlled access layer.

“It supports the docking of various models, allowing for one-click integration across multiple models, with unified protocols, permissions, and disaster recovery.”article, stated

`observability-tracing` (operations)

Otel traces reasoning and retrieval so operators can determine whether failures come from the knowledge base, RAG, or tools.

“we can trace the entire customer search and reasoning processes through an end-to-end tracking system to analyze whether the issue lies with the knowledge base, RAG, or toolset”article, stated

`prompt-versioning` (prompt-ops)

Nacos provides a dynamic prompt update mechanism with monitored gray configuration for post-launch prompt changes.

“through Nacos, we can enable dynamic real-time pushing, obtaining timely optimization effects.”article, stated

`data-quality-loop` (data)

The article treats customer feedback, evaluation data, and data-quality optimization as a flywheel for improving enterprise agent performance.

“Establish a customer feedback and tracking system at the frontend.”article, stated

`rag-retrieval` (retrieval)

RocketMQ and the RAG knowledge base keep system and customer data fresh for inference.

“Use Apache RocketMQ to dynamically update RAG data for real-time feedback and optimization.”article, stated

Tradeoffs

Choose a data-centric multi-agent platform

Chose multi-agent collaboration grounded in private data and feedback loops over fixed single-task agents or waiting for a general AGI agent.

Rationale: Alibaba argues that practical enterprise systems need data quality, tools, and orchestration before AGI-style generality arrives.

“Considering generality and specificity, a balance is often needed, and cost effectiveness must also be balanced.”article, stated

Use gateway-mediated model and tool access

Chose Higress shielding models, tools, permissions, TLS, and quotas over letting each agent integrate directly with every model and tool.

Rationale: The gateway centralizes security and governance while still exposing multiple models and data sources.

“With Higress, we can enable end-to-end TLS on the model access chain to protect data security.”article, stated

Make prompt and RAG data dynamic

Chose Nacos for prompt updates and RocketMQ for data-change synchronization over static prompts and periodically refreshed retrieval data.

Rationale: The system is designed for continuously changing customer and system data.

“The agent comprises a multitude of prompt words and algorithms; through Nacos, we can enable dynamic real-time pushing”article, stated

Gotchas (what bit them)

The article's performance claims are not independently explained.

Root cause: Alibaba reports high resolution rates but does not provide evaluation methodology, sample size, or baseline comparison in the source.

Resolution: Treat the >95% and >85% numbers as vendor-reported outcome claims, not independently validated evidence.

“solving over 95% of consulting issues and over 85% of anomalies.”article, stated

Enterprise agent quality depends on more than the LLM.

Root cause: Retrieval data, tools, prompt words, and customer feedback all change after launch.

Resolution: Build data quality observation, prompt updates, and real-time RAG synchronization into the platform.

“Both system data and customer data are constantly updated.”article, stated

Model access can create security, quota, and key-leakage risks.

Root cause: Agents need access to model providers and internal tools, but direct credentials and unmanaged calls are risky.

Resolution: Centralize model access through Higress with TLS, content safety, API-key management, traffic controls, and quotas.

“we can manage API keys centrally to enhance concurrency, provide internal API keys to agents, prevent API key leakage risks, and control traffic and quotas”article, stated

Evaluation

The source describes observation and vendor-reported production claims but does not provide a rigorous public evaluation design.

Metrics: over 95% of consulting issues solved · over 85% of anomalies solved · Otel tracing of reasoning and recall effectiveness

“Based on the Otel observation system, we can automatically analyze the effectiveness of reasoning processes and recall results.”article, stated