Agentic AI Customer Support Agent on AWS AgentCore

The Challenge: From Demo to Dependable Support

Most AI support demos work until they meet real customers - then they forget who they spoke to yesterday, call tools they should not, or answer confidently with the wrong policy. Production support needs memory that persists across sessions, tool access that is authenticated and governed, and a way to measure whether answers are actually correct.

This project delivers a production-grade customer support agent on Amazon Bedrock AgentCore that handles real e-commerce workflows end-to-end - from product questions and return policy checks to warranty verification and personalized follow-ups for returning customers.

“The agent answers product questions, checks return policies, looks up warranty status via a real Lambda function, searches the web, and remembers returning customers across sessions.”

Real-World Support Workflow

Built for how support actually works - customers ask about products, return windows, and warranty coverage, then expect you to remember them next time. Core capabilities:

Policy-Grounded Answers

get_return_policy and get_product_info tools enforce real business rules, augmented by Exa AI web search for up-to-date context

Customer Memory

Semantic + summarization memory remembers names, preferences, and purchases so returning customers feel recognized

Warranty Verification

A live Lambda function exposed as an MCP tool through AgentCore Gateway for real warranty lookups

Secure Access

Amazon Cognito JWT authentication enforced on both the agent runtime and the gateway

Full Observability

CloudWatch traces, logs, and a GenAI Observability dashboard with X-Ray transaction search for every turn

Quality Evaluation

Continuous LLM-as-Judge scoring on goal success, answer correctness, and tool selection

Customer Chat UI

Flask web interface with Cognito login calling the AgentCore REST API for live conversations

Cedar Guardrails

Policy engine that caps refunds, restricts warranty access by identity, and blocks sensitive-data exposure

Architecture Overview

The system is composed of cooperating layers on AgentCore, designed for production operation:

Agent Layer

Strands Agents on AgentCore Runtime with Amazon Nova Pro. The agent decides when to call custom tools, the web search connector, or the gateway-exposed Lambda tool - no hard-coded call order.

Memory Layer

Shared AgentCore Memory with SEMANTIC and SUMMARIZATION strategies. Sessions are isolated per user ID, but recall works across sessions, so a returning customer is recognized by name and history.

Gateway Layer

AgentCore Gateway exposes the warranty-check Lambda as a strictly-typed MCP tool and manages OAuth egress credentials for external connectors like Exa AI web search.

Governance Layer

A Cedar policy engine gates every tool call: refund amounts are capped, warranty lookups require an authenticated subject from the Cognito JWT, and sensitive-info patterns are blocked.

Evaluation Layer

Online evaluators score live traces with LLM-as-Judge, while a parallel runtime A/B-tests config bundles with AI-driven prompt recommendations before promotion.

Customer support agent architecture with custom tools and Exa AI web search — Fig 2: Agent runtime with custom tools and MCP web search - the entry point for every customer conversation

Declarative Infrastructure

Everything is declared in agentcore.json and deployed through the AgentCore CLI, which synthesizes AWS CDK stacks. Runtimes, memory, gateway, policy engine, harnesses, and evaluation configs are independent resources, so the entire support platform is reproducible in any AWS account from a single deploy:

# Deploy prerequisites (Cognito, Lambdas, SSM params)
aws cloudformation deploy \
  --template-file cloudformation/prereqs.yaml \
  --stack-name agentcore-prereqs \
  --capabilities CAPABILITY_IAM CAPABILITY_NAMED_IAM

# Deploy agent, memory, gateway, policies, harnesses
cd app/CustomerSupport
uv sync
agentcore deploy

# Talk to the live agent
agentcore invoke "What's the return policy for electronics?" --stream

Governing Agent Actions with Cedar

Trust is enforced at the gateway, not left to the model. Raw Cedar policies gate every tool call, so protections hold even under prompt injection:

•Refund limits: refund tool calls above a threshold are denied regardless of what the agent decides
•Warranty access: warranty lookups require an authenticated subject from the Cognito JWT - anonymous callers cannot reach the tool
•Sensitive info protection: prompt-attack and data-exposure guardrails block risky interactions before tools ever run

Cedar policy governance architecture for agent tool access — Fig 3: Cedar policies evaluated at the gateway - every tool call is authorized before it executes

Autonomous Operations & Continuous Improvement

Operational scale is built in. Declarative harness agents extend the platform without additional agent code - with OAuth egress, shell access, and human-in-the-loop approvals wired in. A parallel runtime A/B-tests config bundles, scores both variants with the same LLM-as-Judge evaluators, and promotes the winner back into the declarative config - so the support agent improves without manual redeploys.

Order Research Agent

Autonomous research via gateway tools and shell access, pausing for human approval before acting

Persistent Reporting

Generates and persists reports across sessions with durable storage

Container Runtime

Container build type via CodeBuild to ECR - same declarative model, no custom agent code

A/B Quality Loop

AI-driven prompt and tool-description recommendations tested against live traffic

Efficient by Design

Built for cost-efficient production. The agent runs on amazon.nova-pro-v1:0 - pre-approved in Bedrock, roughly $0.0008 per 1K input tokens, and strong enough for nuanced support reasoning without the cost or approval overhead of larger frontier models. AgentCore Runtime and model invocations are the only billable surfaces; the entire stack tears down cleanly via CDK when not in use.

Skills & Technologies

Amazon Bedrock AgentCoreStrands AgentsAmazon Nova ProMCPAWS LambdaAmazon CognitoCedar PoliciesCloudWatchAWS CDKCloudFormationFlaskPython