Building a Financial‑Grade AI Agent Platform with AgentScope: A Practical Whitepaper

Product Overview

FinXScope is the financial‑grade AI‑native agent base built by Alibaba Cloud's new financial technology service team on top of AgentScope Java. It serves as the core engine of the Agent Harness system, reusing AgentScope’s capabilities for agent orchestration, tool integration, multi‑model access and extensible runtime, and adds deep customizations for compliance, security and high availability.

Since launch, FinXScope has been adopted by 10 leading financial institutions, is in delivery for more than 20 customers and has been deployed for over 60 financial clients, covering state‑owned banks, joint‑stock banks, insurance and securities firms. In production, AI‑native apps built on FinXScope have reached tens of millions of users, proving stability and business value.

Industry Background and Paradigm Shift

Since 2024, autonomous‑agent products such as Manus, Claude Code and Hermes Agent have sparked a paradigm shift: large models evolve from chat assistants to “work partners” that can plan and execute complex tasks. Traditional AI services merely add a “chat” entry to existing workflows, leaving users to navigate menus and forms. The new “conversation‑as‑a‑service” model lets users describe goals in natural language and lets the AI autonomously plan, schedule and invoke hundreds of skills.

Six‑Layer AI‑Native Architecture

FinXScope defines a six‑layer architecture that replaces the traditional SOA/Microservice stack with an AI‑centric stack:

Intent Engine : Four‑stage pipeline (NER, query rewriting, intent classification, skill mapping) with configurable modules and LLM fallback.

Three‑Tier Memory : Short‑term memory (STM) with sliding‑window token compression, long‑term memory (LTM) stored in PostgreSQL + pgvector, and UserProfile built from LTM data.

Dual‑Mode Execution : OneAgent (ReAct loop) for autonomous planning and Multi‑Agent with eight orchestration strategies (Parallel, Sequential, Loop, MsgHub, Debate, Graph, Routing, Supervisor).

AG‑UI Protocol : Unified event stream for synchronous, streaming and UI interactions, supporting cross‑channel consistency.

Three‑Level Skill Definition : YAML for simple queries, SKILL.md + script for medium complexity, Java @Bean for complex business logic.

Tool & Knowledge Integration : MCP and API‑Schema connectors, unified Knowledge interface for RAG, domain‑specific KBs, and permission propagation.

Security, Compliance and Observability

Three‑Level Permission : User‑Agent‑Skill checks with FILTER/REJECT modes and configurable caching.

Input Protection : Prompt‑injection defense, sensitive‑word filtering and content audit.

Full‑Trace Auditing : Structured logs with traceId/spanId, complete decision‑chain records.

Observability Stack : Micrometer tracing, Prometheus metrics exposed at /actuator/prometheus, BizLogger dual‑mode output, SPI hooks for custom metrics.

Deployment and Evolution

FinXScope can be packaged as a standalone Fat JAR/Docker image or used as a Spring Boot starter. Version 2.0 adds a multi‑agent workflow engine, configuration hot‑update via Redis Pub/Sub, stateless STM/LTM, and a full‑stack operations platform.

Typical integration steps: (1) import the starter or run the container, (2) configure intent trees, skills and tools, (3) enable AG‑UI, (4) launch the service. Resource estimates assume reuse of existing PostgreSQL, Redis and OSS; only missing components are deployed.

Typical Scenarios

Wealth‑management chat, client‑manager assistance, intelligent underwriting, compliance review, investment research.

Long‑running tasks such as cross‑border remittance or loan approval are managed by the TodoList subsystem, which provides real‑time progress, planning validation and checkpoint recovery.

Conclusion

FinXScope demonstrates that a financial‑grade AI agent base can be built on top of an open‑source framework, delivering a complete stack from low‑code prototyping to high‑code production, while meeting the high‑availability, security and observability requirements of the finance industry.