CTO · VP Engineering · Agentic Architecture

The agentic architecture
your engineers will ask
about before they build.

Production-grade agentic architecture for enterprise operational workflows requires three capabilities that most agent frameworks do not provide: durable state that persists across multi-week wait states, machine-to-machine coordination via open protocols, and AI client integration that works with any model or tool.
The engineering choices that distinguish PLRX from frameworks and workflow tools are architectural, not configuration — durable execution, exactly-once semantics, and open protocol interoperability are designed into the substrate, not added as middleware.
This page covers the three architectural pillars behind every PLRX mission — for engineering leaders evaluating whether to build this capability or deploy it.

94% autonomous resolutionFrom $0.99 per missionEnterprise AgenticA2A · MCP · Open protocols

Book a Scoping Call

Discuss the architecture.

Tell us which operational workflow your engineering team is evaluating for agentic deployment. Proof of concept in 2–3 weeks — production in 12 weeks.

First Name Required.

Last Name Required.

Work Email Please enter your corporate email address.

Company Required.

Job Title Required.

Something went wrong. Please try again or email [email protected].

By submitting you agree to our Privacy Policy. We never sell your data.

Request received.

Someone from the PLRX team will be in touch within two business days to schedule your scoping call.

Three Architectural Pillars

Every PLRX mission runs on three architectural layers. Each addresses a specific production engineering requirement that generic frameworks leave to the application developer.

Pillar 01 · Durable State Machine

Workflows that survive everything — weeks, crashes, and network failures

Every mission runs inside the PLRX Durable State Machine. Workflow state is persisted to an append-only event log. On worker restart, workflows replay deterministically and resume at the exact last committed state. Missions pause in input_required state for minutes, hours, days, or weeks — waiting for a provider document, a payer callback, a supplier confirmation — without consuming resources or losing state. Exactly-once semantics on every side-effecting operation: deterministic idempotency keys prevent duplicate submissions, emails, or API writes on retry. This is not middleware — it is the substrate every agent mission runs on.

Pillar 02 · A2A Protocol

Open machine-to-machine coordination across agent boundaries

PLRX implements the Agent-to-Agent (A2A) protocol — an open JSON-RPC 2.0 standard for machine-to-machine coordination. Specialist agents coordinate via A2A: an orchestrator mission delegates to a Clinical Documentation Specialist, a Billing Service Agent, or a Data Scientist Agent via structured A2A message/send calls. Each specialist runs in its own execution context, returns results to the orchestrator, and the mission continues. A2A is inspectable, composable, and portable — your systems, your integrations, and third-party agents can participate in the same protocol.

Pillar 03 · MCP Server

Universal AI client integration — any model, any tool

The PLRX MCP (Model Context Protocol) Server provides a universal integration layer for AI client tools — Claude, Cursor, any MCP-compatible client — to connect to the PLRX platform and interact with live operational data, mission state, and agent capabilities. MCP is an open standard: the same integration point works for any AI client tool that speaks the protocol, without bespoke per-tool integrations. For engineering teams building internal AI tooling on top of production operational data, MCP is the architectural interface.

Architecture Reference — What Each Layer Provides

The technical architecture
behind every PLRX mission.

Component	What It Provides	Engineering Significance
Durable State Machine	Persistent workflow execution with exactly-once semantics. State stored in append-only event log. Deterministic replay on restart. Input_required suspension without resource consumption. Configurable retry policies with exponential backoff.	Eliminates the stateless execution problem that breaks agent frameworks on multi-week operational workflows. Engineers do not write state persistence code — the platform guarantees it.
A2A Protocol Server and Client	JSON-RPC 2.0 machine-to-machine coordination. Task management API with streaming, push notification, and polling modes. Specialist agent delegation and result aggregation. Open standard — portable to any A2A-compatible agent.	Provides the coordination primitive for multi-agent missions without proprietary lock-in. The same coordination protocol works for PLRX specialist agents and customer-built agents operating on the same platform.
MCP Server	Universal AI client integration gateway. Exposes PLRX operational capabilities, mission state, and agent actions to any MCP-compatible AI client. Structured tool definitions for mission management, state inspection, and agent invocation.	Enables AI-powered internal tooling — dashboards, copilots, investigative tools — to operate on live production operational data through a standard interface, without bespoke per-tool integrations.
Kubernetes on AWS	Multi-tenant or dedicated deployment on AWS Kubernetes. Sovereign per-tenant environments with no shared runtime or data plane. Horizontal scaling at the infrastructure layer without application-level configuration changes.	Production infrastructure for enterprise workloads. Tenant isolation is architectural — separate Kubernetes environments, not logical partitions. PHI and sensitive data never traverse shared infrastructure.
WORM Audit Trail	Write-Once Read-Many event log capturing every agent action — model, timestamp, input, output, decision, state transition. Append-only, object-locked, queryable without platform access.	Compliance-grade record for regulated industry deployment. Not a logging feature — the same event log that provides exactly-once replay also provides the complete audit trail. Architecture and compliance are the same layer.

CTO · The Architecture Question That Determines Build vs Deploy

How long does it take to build this from scratch — and what does it actually require?

Building production-grade agentic architecture for enterprise operational workflows requires five concurrent specialisations: durable workflow engineering (determinism, fault tolerance, idempotency), vertical domain expertise (payer protocols, compliance rules, exception patterns), production ML (model selection, evaluation, inference pipeline design), compliance architecture (WORM audit logs, PHI encryption, exactly-once guarantees for regulated operations), and platform DevOps (Kubernetes, multi-tenant isolation, observability).

Teams that have assembled all five estimate 12–15 months to first reliable production mission. Without prior durable workflow experience, longer. The engineering challenges are well-defined — deterministic replay, idempotency at every external boundary, suspension patterns for week-long waits — but each one requires specific expertise to implement correctly under production conditions.

PLRX is the alternative: deploy the pre-built architecture and the pre-built specialist agents in 12 weeks. Your engineering team contributes to integration, configuration, and governance — not to re-implementing production infrastructure that already exists.