System Architecture

Architecture: Protocol Stack & Agent Topology

ARCNAT is built on a protocol-native foundation where 80+ intelligent agents coordinate across 20 waves through 8 interoperability protocols. This is not AI bolted onto a legacy system — it is an autonomous operating system designed from the ground up for retirement plan administration.

Full Technical Documentation Request Architecture Brief

Protocol Stack

Eight interoperability protocols govern how agents discover, communicate, and execute across the system.

MCP (Model Context Protocol)

Universal data and tool bus. Every data source, calculation engine, and external API is exposed as an MCP server.

A2A (Agent-to-Agent)

Structured inter-agent communication. Every cross-agent interaction uses capability matching and response acknowledgment.

ANP (Agent Network Protocol)

Agent registry, discovery, and health monitoring. Load distribution and cross-tenant isolation.

AG-UI (Agent-User Interaction)

Real-time agent activity streaming via WebSockets and SSE. Users see agents work, never stare at a spinner.

AGORA

Natural language orchestration. Staff can instruct the system in plain English.

Eclipse LMOS

Transport-agnostic integration layer. REST, SFTP, SOAP, WebSocket, MQTT — one agent logic, different adapters.

AP2 (Agent Payment Protocol)

Autonomous financial transaction execution. Bounded by authorization limits and fiduciary rules.

OASF (Open Agentic Schema)

Agent capability publication. Every agent publishes its "resume" for discovery and marketplace listing.

Agent Topology

80+ intelligent agents are categorized across 5 tiers — orchestration, domain, sentinel, governance, and intelligence — spanning 20 waves.

Orchestration — workflow and task coordination

Census & Data — participant and plan data

Eligibility & Provisions — entry and benefit rules

Compliance Testing — ADP, top-heavy, 410(b), etc.

Government Filings — 5500, 1099-R, DOL/IRS

Trust Accounting — contributions and reconciliation

Distributions — withdrawals and rollovers

Plan Documents — SPDs, amendments, restatements

Corrections — EPCRS and remediation

CRM & Sales — contacts and pipeline

Projects & Workflows — tasks and deadlines

Financial Management — billing and invoicing

Intelligence & Knowledge — RAG, analytics, insights

Platform Governance — security, policy, admin

The 7-Pillar Architecture

Seven foundational pillars that ensure ARCNAT operates with deterministic proof, vendor neutrality, and long-term survivability.

1. Canonical Spine

Single source of truth for all data

Complete data model with 20+ entity categories

Tenant isolation at database level (RLS)

Event sourcing with outbox pattern

Digital twin for plan state representation

2. Event Backbone

Kafka-driven event streaming architecture

Immutable event log for audit trail

Outbox pattern for reliable delivery

Event schema registry

Workflow event correlation

3. Policy Engine

Single canonical policy-as-code system

ABAC (Attribute-Based Access Control)

Entitlement management

Tier enforcement

Hard-stop policy gates

4. Agent Runtime

MCP-gated AI agents with proof-gated orchestration

80+ intelligent AI agents

Model Context Protocol (MCP) gates

Multi-agent collaboration

Proof-gated orchestration

5. Experience Surfaces

Multiple interfaces for different user types

Internal UI (Next.js App Router)

PSL (Plan Sponsor Link) portal

GraphQL API with subscriptions

REST API v1

6. Ops Intelligence

Full observability and monitoring

OpenTelemetry tracing

Structured logging

Metrics collection

Agent evaluation and monitoring

7. Platform Reliability

Multi-region disaster recovery and graceful degradation

Multi-region disaster recovery ready

Health check endpoints

Graceful degradation

Cost-efficient design

Core Architectural Principles

The design decisions that make ARCNAT suitable for regulated environments.

Evidence-by-Default

Every decision produces proof. Every workflow step generates evidence. Every agent action is logged. Evidence is bundled automatically, sealed, and export-ready for audits, litigation, and regulatory defense.

Immutable event logs
Automatic evidence bundling
Legal hold and redaction support
Deterministic replay capability

Vendor Neutrality

Vendors are adapters. The kernel never changes. When you switch vendors, integrations, or AI models, your evidence, workflows, and data remain intact. The architecture survives vendor changes.

Model/provider abstraction layer
Adapter pattern for integrations
Portable evidence format
Kernel-vendor separation

Deterministic Replay

Reconstruct any outcome from the past. Given inputs, rules, approvals, and timestamps, ARCNAT can replay any workflow deterministically. Essential for audits, litigation, and proving correctness years later.

Event-sourced architecture
Immutable state snapshots
Rule versioning
Complete audit trail

Canonical Truth

One data spine. One workflow truth. One evidence truth. No synchronization issues. No data conflicts. No ambiguity about what happened or why.

Single source of truth
Event-driven state updates
Digital twin representation
Consistent data model

AI & Agent Architecture

Proof-gated orchestration ensures AI autonomy is compatible with regulated environments.

Proof-Gated Orchestration

AI proposes, ARCNAT proves and enforces

Deterministic Calculation Services

Source of truth - LLM cannot be the calculator. All financial calculations use deterministic services.

Hard-Stop Policy Gates

Can block execution - cannot be bypassed. Policy enforcement happens before execution.

Event-Native Evidence

Automatically collected from event streams. No manual evidence gathering required.

Model/Provider Abstraction

Vendor-neutral, portable evidence. Switch AI models without losing evidence.

80+ Intelligent AI Agents

Each agent has a specific responsibility:

Workflow orchestration agents
Compliance testing agents
Document processing agents
Data extraction agents
Risk assessment agents
And 24 more specialized agents

Multi-Model Support

18+ LLM models supported:

GPT-5.2, Claude 4.5, Gemini 2.5
17 reasoning patterns
Advanced RAG systems
Knowledge graphs
Predictive analytics
Model Context Protocol (MCP)

Security & Compliance Architecture

Built for environments with zero tolerance for failure.

Zero-Trust Security Model

Every access verified, every action logged

RBAC / ABAC enforcement

End-to-end encryption

Row-level security (RLS) for tenant isolation

Compliance-Ready Exports

Evidence bundles (sealed and export-ready)

Audit reports with deterministic replay

Regulatory filings (formatted and sealed)

Legal hold and redaction support

Explore the Full Architecture

For detailed technical documentation, API references, database schemas, and implementation guides, see the complete architecture documentation.

View Full Documentation Request Architecture Brief

Solutions · Security · Compliance · Production Readiness