AWS Architecture Blog·August 20, 2026Building Production-Grade AI Sales Agents: Architecture and Pillars
This article details AgentFlo's architecture for building scalable, production-grade AI sales agents using Amazon Bedrock AgentCore. It outlines the five pillars of Velocity, Standardization, and Scalability (with Trust and Reliability in Part 2) that guide their system design. The core system involves orchestrating LLMs with various tools, context, and memory components to handle complex sales conversations and integrate with commerce systems.
Read original on AWS Architecture BlogAgentFlo developed an architecture for intelligent sales agents leveraging Amazon Bedrock AgentCore and the Strands Agents SDK to address challenges like high cart abandonment and the inability of traditional chatbots or human agents to scale personalized sales interactions. The system aims to convert conversations into completed purchases across various messaging channels.
Core Components of an AI Agent
An AI agent, in AgentFlo's context, is a sophisticated system integrating several key components to deliver intelligent conversational capabilities. Understanding these components is fundamental to designing robust AI-driven applications:
| Component | Function in AgentFlo |
|---|
- Model: The Large Language Model (LLM) at the core, responsible for understanding user intent and determining the next action.
- System Prompt / Persona: Defines the agent's role, tone, and behavioral limits (e.g., sales agent, support agent).
- Tools: External functions or APIs (e.g., product search, inventory check, cart creation, order placement) that allow the agent to perform actions in the real world.
- Knowledge: Merchant-specific data sources (product catalogs, policies, FAQs) used to ground responses and provide accurate information.
- Memory / State: Stores conversation history, cart state, customer preferences, and previous actions to maintain context and coherence across interactions.
- Channels: Integrates the agent with various communication platforms like WhatsApp, SMS, web chat.
- Guardrails / Policy: Mechanisms to prevent unsafe, unauthorized, or incorrect actions and to filter harmful content.
- Observability: Systems for tracking performance, costs, conversions, and conversation quality.
AgentFlo Production Architecture on AWS
AgentFlo's architecture on AWS is designed to be highly scalable, secure, and performant. Messages from customer channels like WhatsApp flow through an Application Load Balancer to an AWS Fargate-based messaging layer. This layer handles initial processing, including authentication, OCR, speech-to-text, and guardrails like prompt injection detection.
Validated requests then enter the AgentCore runtime, a capability of Amazon Bedrock AgentCore, orchestrated by the Strands Agents SDK. This runtime manages individual agent sessions in lightweight virtual machines, ensuring tenant isolation and security. The AgentCore Gateway brokers tool calls to an API layer built with AWS Lambda functions (for Cart, Product, Knowledge Base services). State is persisted across Amazon DynamoDB (sessions, carts), Amazon Aurora (orders), and an Amazon Bedrock Knowledge Base backed by Amazon S3. Observability is handled by AgentCore Observability and Amazon Data Firehose, which captures interactions for analytics in Amazon S3.
Key Architectural Decisions for AI Agents
AgentFlo's choice of Amazon Bedrock AgentCore was driven by its support for stateful sessions for long-running conversations, a secure agent runtime with hardware-level security boundaries per tenant using lightweight VMs, and native Model Context Protocol (MCP) integration for standardized, authorized tool connectivity. These features are critical for building reliable and secure multi-tenant AI agent platforms.