We didn’t start by asking how to add agents to underwriting. We asked how underwriting should work in an agentic world, and built from there.
After months of quiet iteration, we're opening the doors to what we've been building:
The Agentic Underwriter Workbench a modern, extensible platform designed from first principles to support experimentation, scale, and AI-native underwriting.
Demo at Future of Agentic Underwriting
Why We Built This
Too many GenAI projects start with the tech and retrofit the workflow.
We did the opposite. We mapped how underwriting should work if built today with connected agents, schema-first design, flexible product modeling, and an interface that treats intelligence as a teammate, not a sidecar.
This is the foundation.
What Makes It Different
Here’s a preview of the architectural highlights:
Cortex: Pluggable AI Engine
Built in FastAPI with swappable LLMs (OpenAI, Anthropic, Gemini, Phi) via config, Cortex is designed for goal-based, multi-agent orchestration and modular tool invocation via JSON-RPC.
Schema-Driven Dynamic Forms
Using rjsf + shared AJV-validated datacontracts, we render multi-product forms (Life, Auto, Renters...) from JSON Schema. The data capture form for product lines are defined at the schema level to avoid costly UI rewrites needed for new product lines.
Activity Stream: A Living Timeline
Think GitHub meets audit logs. Built with Zustand, Fuse.js, and UI card components, the stream captures all events (ratings, comments, file uploads) in a searchable, real-time feed.
Agentic UI
The assistant panel is persistent, context-aware, file-friendly, and can run in recommend or autonomous mode, just like a teammate. It integrates deeply with the rest of the workflow, not bolted onto the side.
Design Language Inspired by NotebookLM
We were inspired by what made Google’s NotebookLM intuitive and borrowed its principles:
Left nav for queue context
Central workspace for focus
Right assistant panel for support
What's Next
This post lays the groundwork. In future installments, we’ll go deeper into:
Cortex’s evolving multi-agent orchestration engine
The JSON-RPC + MCP architecture behind tool invocation
UI design trade-offs around schema-first workflows
Creating agent memory and semantic retrieval layers
Embedding underwriting intelligence into every pixel of the platform
We’re building this in the open.
If you're working on underwriting, claims automation, agentic workflows, or GenAI-driven UIs—we’d love to connect.
Built by a folks that’s lived the technology pain in insurance, and dont want to duct-tape AI to legacy systems.
Feel free to check us out at http://helloinsurance.substack.com