Shawn Xie

Freyja

AI Native Intelligent Insurance Configuration

Multi-agent relay—taking on complex insurance configuration

Multi-AgentPlanningMCPHuman-in-the-loop
Unmanned shipyard assembly—modular hull blocks, robots and AGVs, sea opening—build-phase metaphor for the agent team

Value

Core value of this project

Shorter end-to-end configuration path

B2B teams submit requests in natural language; multi-agent orchestration shapes requirements, configures modules, and coordinates release—reducing the long chain of “standardize the doc first, then click through low-code.” Human confirmation at key gates.

Surface information gaps and platform limits early

Before large-scale configuration, unified validation and interactive confirmation separate executable items, missing information, and platform constraints—lowering late rework and repeated alignment cost.

Orchestrate cross-system, cross-module configuration and release

One task advances products, rules, documents, portals, and related modules by dependency, including environment release—avoiding missed cross-module steps and ordering errors.

Architecture

System architecture and configuration workflow

Multi-Agent Architecture

  • Layered delegation: unified orchestration → domain sub-agents → diverse tool calls.
  • Structured handoff: a JSON requirement brief as the collaboration artifact—less information loss between users, the root agent, and sub-agents.
  • Aligned with the existing platform’s configuration and runtime layers—not a parallel business stack.

Configure-to-Release Flow

  • Intent → unified validation → human confirmation → dynamic planning → modular execution → self-check → release.
  • Scope and module order are orchestrated on demand—not a fixed template walked step by step.
  • Exception paths: automatic degradation and engineer escalation.
  • For quality evaluation, regression, and release gates, see Freyja Eval

Outcomes

Results and evidence

Links