A multi-agent AI system was built to automate reinsurance data processing using Temporal workflows. The system extracts catastrophe loss data from non-standardized Excel submission packs, matches events to historical records, and populates internal databases. Key architectural decisions include breaking the process into modular sub-agents (Submission Pack Parser, Historical Matcher, etc.) to maintain focus and reduce LLM confusion, implementing human-in-the-loop confirmations for every tool execution and agent completion using Temporal Signals, and using a Bridge Workflow to route interactions and store inter-agent data. The approach prioritizes reliability and human oversight over full automation, treating AI agents as supervised assistants rather than autonomous systems.
Table of contents
Reinsurance 101 #The business problem #Agents and tools #Why agentic AI? #Why multiple agents? #Implementing an agent with human-in-the-loop #Multiple agents with human-in-the-loop #Principles for building AI agents #Conclusion #Sort: