How Ant Financial’s Multi‑Agent Platform “Tiangong Wànxiàng” Transforms Front‑End AI Automation

Introduction

The emergence of AutoGPT in 2023 marked a new stage for agents, combining GPT‑3.5 and GPT‑4.0 to handle complex project tasks, yet it was still limited by base model capabilities and token constraints. The release of Manus on March 6, 2023 demonstrated a dramatic leap in autonomous decision‑making and tool‑calling abilities.

Core Concepts

LangGraph

LangGraph replaces the traditional linear chain with a directed‑graph model, enabling non‑linear AI workflows. Nodes and edges support loops, branches, and parallel processing, allowing dynamic tool‑calling paths such as "agent → tools → agent" that linear chains cannot achieve.

Multi‑Agent System (MAS)

A Multi‑Agent System consists of multiple autonomous agents that collaborate, compete, or negotiate to accomplish complex tasks. Each agent has independent decision‑making, can share context via a global checkpointer, and the system distributes work across agents to reduce overall complexity.

Tiangong Wànxiàng System Design

Architecture Core Idea

The platform is built as a flexible, extensible architecture that can easily integrate new sub‑agents (MCP) and handle both vertical business logic and generic tasks.

Technical Choices and Rationale

Why a Multi‑Agent Architecture Instead of a Unified Agent

Unified agents like Manus rely on a centralized scheduler (a dynamic To‑Do list) which becomes a bottleneck when task types span large contexts. Multi‑Agent design offers inherent context advantages, better scalability, and easier extension by adding or swapping agents.

Why ReAct Paradigm Instead of Fixed Workflows

ReAct (Reasoning‑Action‑Observation) loops enable dynamic decision‑making and on‑demand tool invocation, which static workflows (e.g., predefined LLM → tool chains) cannot provide. This is crucial for high‑accuracy scenarios such as code generation and data analysis.

{
  name: 'sequentialThinking',
  description: `一个用于动态和反思性解决问题的详细工具。`
  // ... (rest of the JSON definition as in the source)
}

Context Engineering

Tiangong Wànxiàng implements a global checkpointer for cross‑agent memory sharing and extends LangGraph’s MemorySaver with a Zcache‑based context manager, enabling session‑level memory sharing across data centers for up to one hour.

Professional Agent Construction

Web‑Page Development Expert

Supports two generation modes: (1) React‑based middle‑back‑office pages that can call real back‑end APIs and be debugged within the platform, and (2) pure static pages that require no API calls and can be shared via a generated URL. Example outputs include CRUD admin pages, data‑dashboard visualizations, TPS algorithm visualizations, and Mermaid diagram rendering.

Same‑Industry Agent "Xiao Qiu"

Built on a natural‑language data‑retrieval agent, it adds network search, context memory, and streaming pipelines to provide data‑analysis capabilities.

All‑Purpose Assistant

Aggregates capabilities from the code‑generation and same‑industry agents (search, image generation, planning tools) to serve as a general‑purpose conversational assistant, comparable to DeepSeek or Baidu’s BaLing.

Conclusion

The article shares the team’s reflections and experiences in building Tiangong Wànxiàng, inviting interested readers to discuss and collaborate.