AI Agents

OpenInsure runs AI workloads as Cloudflare Durable Objects backed by the agents npm package (^0.7.5). Each agent is a stateful, long-lived object that can hold WebSocket connections, maintain conversation history, and coordinate with other agents via the OrchestratorAgent.

Architecture

POST /v1/... or WebSocket
        │
        ▼
  OrchestratorAgent (DO)
        │
   ┌────┴──────────────────────────────────┐
   │                                       │
   ▼                                       ▼
SubmissionAgent   UnderwritingAgent   ClaimAgent  ...
   │                    │                  │
   ▼                    ▼                  ▼
EmailIntakeAgent  WorldIntelligence   AdjudicationAgent

All agents share:

Model: Claude (claude-sonnet-4-6 for reasoning, claude-haiku-4-5 for fast extraction)
Observability: Langfuse for tracing every prompt, response, tool call, and latency
State: Durable Object SQLite for conversation history and task state

Agent Inventory

Durable Object	Purpose
`SubmissionAgent`	Processes new business submissions end-to-end
`PolicyAgent`	Mid-term policy actions (endorsements, cancellations)
`ClaimAgent`	Claims workflow automation and AI-assisted investigation
`UnderwritingAgent`	Risk assessment, rate justification, referral drafting
`AdjudicationAgent`	Coverage determination and denial letter drafting
`BordereauxAgent`	Automated bordereaux generation and carrier submission
`ComplianceAgent`	Regulatory deadline tracking and filing automation
`ProducerAgent`	Producer onboarding and licensing workflow
`TelematicsAgent`	Telematics data ingestion and risk scoring
`LossControlAgent`	Loss prevention recommendations
`WorldIntelligenceAgent`	External risk data enrichment (weather, FMCSA, news)
`EmailIntakeAgent`	Shared mailbox parsing and routing
`NotificationAgent`	Intelligent notification dispatch and scheduling
`CollaborationAgent`	Deal Room real-time collaboration
`AssemblyAgent`	Document assembly and form population
`OrchestratorAgent`	Multi-agent task coordination
`OpenInsureMCP`	MCP server for external AI integrations
`DealRoomDO`	WebSocket-backed carrier deal room

Submission Automation

The SubmissionAgent handles the full intake-to-quote workflow when triggered by either the producer portal or the email ingest cron.

Email intake flow

EmailIntakeAgent receives a new message from Underwriting@mhcmga.com via Graph API 2. Classifies the email (new submission, renewal, endorsement request, general inquiry) 3. Extracts attachments and stores them in R2 4. Invokes AssemblyAgent to extract structured data from ACORD forms using Workers AI + Claude 5. Creates a draft submission record in PlanetScale with extracted fields 6. Notifies the UW queue via the Notification system

Risk narrative generation

For every new submission, UnderwritingAgent generates an AI risk narrative using Claude:

// Simplified — actual implementation in packages/agents
const narrative = await agent.generateRiskNarrative({
  submission,
  lossHistory,
  industryBenchmarks,
});
// Output: structured risk summary with flagged items,
// recommended premium adjustment (-5% to +5%), and
// adverse action notice if adjustment is negative

The narrative is attached to the submission record and displayed in the Underwriting Workbench.

Claims AI

ClaimAgent provides AI assistance throughout the claims lifecycle:

Feature	Description
Coverage determination	Analyzes the policy form and loss facts, drafts a coverage opinion
Reserve estimation	Suggests initial reserve based on loss type, jurisdiction, and injury severity
Fraud signal detection	Flags patterns from the loss description and claimant history
Settlement recommendation	Benchmarks against comparable closed claims
Denial letter drafting	Generates state-compliant denial letters from coverage position

World Intelligence

WorldIntelligenceAgent enriches submissions and claims with external data:

FMCSA — Motor carrier safety data for commercial auto and transportation risks
Weather — Storm and catastrophe data for property claims (NOAA integration)
News — Adverse media scan for named insureds on large commercial submissions
NCCI — Workers’ Compensation experience modification factor lookup
ISO — PPC (Public Protection Classification) for property risks

Observability

All agent activity is traced in Langfuse (us.cloud.langfuse.com). Every trace includes:

Agent type and DO ID
Input prompt (sanitized — no PII)
Model used and token counts
Tool calls and results
Latency per step
Submission / policy / claim ID for correlation

Set LANGFUSE_PUBLIC_KEY and LANGFUSE_SECRET_KEY via wrangler secret put to enable tracing.

Model Selection

Use case	Model	Reason
Risk narrative generation	`claude-sonnet-4-6`	Complex reasoning, long context
Form field extraction	`claude-haiku-4-5`	Fast, cost-efficient for structured extraction
Coverage determination	`claude-sonnet-4-6`	Legal reasoning required
Email classification	Workers AI (Llama 3)	Low latency, no external API call

Invoking Agents via API

Internal services invoke agents through the /internal/ prefix, which is auth-gated by API_SECRET:

POST /internal/agents/submission
Authorization: Bearer <API_SECRET>
Content-Type: application/json

{
  "submissionId": "sub_01J8...",
  "action": "generate_narrative"
}

The MCP server (/mcp) exposes agent capabilities to external Claude integrations — see MCP Server for details.