AWS Bedrock AgentCore — Serverless Platform for Production AI Agents

Posted on: 4/25/2026 1:13:40 AM

Building an AI Agent that works in a is one thing. Deploying that agent to production with millions of requests, enterprise-grade security, cross-session memory, and integration with existing business systems — that's the real challenge. Amazon Bedrock AgentCore was built to solve exactly this: a serverless platform providing every building block needed to run AI Agents at production scale.

1. Why Do We Need a Dedicated Platform for AI Agents?

When building AI Agents, the "smart" part — calling LLMs, parsing results, invoking tools — typically accounts for about 20% of the effort. The remaining 80% is production concerns: security, authentication, memory management, monitoring, retry logic, rate limiting, and deployment at scale. This is why managed platforms like Bedrock AgentCore are becoming increasingly critical.

graph TB
    subgraph "AI Agent Development Effort"
        A["20% — Agent Logic
LLM, Tools, Prompts"]
        B["80% — Production Infrastructure
Security, Memory, Observability,
Scaling, Identity, Deployment"]
    end

    style A fill:#4CAF50,stroke:#fff,color:#fff
    style B fill:#e94560,stroke:#fff,color:#fff

2. Amazon Bedrock AgentCore Architecture Overview

AgentCore isn't a single service — it's a suite of modular services, each addressing a specific aspect of the production AI Agent problem. You can use the entire suite or just the parts you need.

graph TB
    DEV["Developer / Agent Framework"]

    subgraph "Amazon Bedrock AgentCore"
        RT["Runtime
Serverless Execution"]
        GW["Gateway
Tool Access & Auth"]
        MEM["Memory
Session & Long-term"]
        ID["Identity
Agent Authentication"]
        OBS["Observability
Tracing & Metrics"]
        EVAL["Evaluations
Quality Scoring"]
        POL["Policy
Action Control"]
        BR["Browser
Web Interaction"]
        CI["Code Interpreter
Code Execution"]
    end

    DEV --> RT
    RT --> GW
    RT --> MEM
    RT --> ID
    RT --> OBS
    RT --> EVAL
    RT --> POL
    RT --> BR
    RT --> CI

    style DEV fill:#e94560,stroke:#fff,color:#fff
    style RT fill:#2c3e50,stroke:#fff,color:#fff
    style GW fill:#16213e,stroke:#fff,color:#fff
    style MEM fill:#16213e,stroke:#fff,color:#fff
    style ID fill:#16213e,stroke:#fff,color:#fff
    style OBS fill:#16213e,stroke:#fff,color:#fff
    style EVAL fill:#16213e,stroke:#fff,color:#fff
    style POL fill:#16213e,stroke:#fff,color:#fff
    style BR fill:#16213e,stroke:#fff,color:#fff
    style CI fill:#16213e,stroke:#fff,color:#fff

3. Runtime — Deploy Agents Without Infrastructure

AgentCore Runtime is the heart of the entire platform. It provides a serverless environment with session isolation — each conversation runs in a separate container, sharing no state with other sessions.

# Example: Deploy agent with AgentCore CLI
# Step 1: Initialize project
agentcore init my-support-agent --framework strands

# Step 2: Define agent logic
# agent.py using any framework (Strands, LangGraph, CrewAI...)

# Step 3: Deploy to AgentCore Runtime
agentcore deploy --name my-support-agent \
    --memory enabled \
    --guardrails my-guardrail-id \
    --identity-provider okta

The biggest differentiator of AgentCore Runtime: it's framework-agnostic. You can use any agent framework — Strands Agents SDK, LangGraph, CrewAI, AutoGen — and Runtime handles deployment, scaling, and monitoring.

4. Knowledge Base — Fully Managed RAG

Amazon Bedrock Knowledge Base provides a complete RAG (Retrieval-Augmented Generation) pipeline: from data ingestion, chunking, embedding, to vector search — all fully managed.

graph LR
    S3["S3 Bucket
Documents"] --> CHUNK["Auto Chunking
Semantic / Fixed"]
    CHUNK --> EMB["Embedding Model
Titan / Cohere"]
    EMB --> VS["Vector Store
OpenSearch / Pinecone"]

    Q["User Query"] --> AGENT["Bedrock Agent"]
    AGENT --> VS
    VS --> CTX["Retrieved Context"]
    CTX --> AGENT
    AGENT --> R["Grounded Response"]

    style S3 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style AGENT fill:#e94560,stroke:#fff,color:#fff
    style VS fill:#2c3e50,stroke:#fff,color:#fff
    style R fill:#4CAF50,stroke:#fff,color:#fff
    style CHUNK fill:#16213e,stroke:#fff,color:#fff
    style EMB fill:#16213e,stroke:#fff,color:#fff
    style Q fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style CTX fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50

Chunking Strategies

5. Guardrails — Output Control and Security

Guardrails are the most critical defense layer in production. Amazon Bedrock Guardrails provides 6 safeguard policies to control both agent input and output.

graph LR
    INPUT["User Input"] --> G1["Content Filter"]
    G1 --> G2["Denied Topics"]
    G2 --> G3["Word Filter"]
    G3 --> G4["PII Redaction"]
    G4 --> G5["Prompt Attack Detection"]
    G5 --> LLM["LLM Processing"]
    LLM --> G6["Contextual Grounding"]
    G6 --> OUTPUT["Safe Response"]

    style INPUT fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style LLM fill:#2c3e50,stroke:#fff,color:#fff
    style OUTPUT fill:#4CAF50,stroke:#fff,color:#fff
    style G1 fill:#e94560,stroke:#fff,color:#fff
    style G2 fill:#e94560,stroke:#fff,color:#fff
    style G3 fill:#e94560,stroke:#fff,color:#fff
    style G4 fill:#e94560,stroke:#fff,color:#fff
    style G5 fill:#e94560,stroke:#fff,color:#fff
    style G6 fill:#e94560,stroke:#fff,color:#fff

Guardrails Configuration Example

{
  "name": "production-guardrail",
  "contentPolicyConfig": {
    "filtersConfig": [
      { "type": "SEXUAL", "inputStrength": "HIGH", "outputStrength": "HIGH" },
      { "type": "VIOLENCE", "inputStrength": "HIGH", "outputStrength": "HIGH" },
      { "type": "HATE", "inputStrength": "HIGH", "outputStrength": "HIGH" },
      { "type": "INSULTS", "inputStrength": "MEDIUM", "outputStrength": "HIGH" }
    ]
  },
  "topicPolicyConfig": {
    "topicsConfig": [
      {
        "name": "competitor-comparison",
        "definition": "Questions comparing products with competitors",
        "type": "DENY"
      }
    ]
  },
  "sensitiveInformationPolicyConfig": {
    "piiEntitiesConfig": [
      { "type": "EMAIL", "action": "ANONYMIZE" },
      { "type": "PHONE", "action": "ANONYMIZE" },
      { "type": "CREDIT_DEBIT_CARD_NUMBER", "action": "BLOCK" }
    ]
  },
  "contextualGroundingPolicyConfig": {
    "filtersConfig": [
      { "type": "GROUNDING", "threshold": 0.7 },
      { "type": "RELEVANCE", "threshold": 0.7 }
    ]
  }
}

6. Multi-Agent Collaboration — Specialized Agent Teams

When a single agent can't handle complex tasks, Bedrock lets you build a team of agents that coordinate with each other. The supervisor-collaborator architecture supports up to 10 collaborator agents per supervisor.

graph TB
    USER["User Request"] --> SUP["Supervisor Agent
Orchestrate & Synthesize"]

    SUP --> A1["Agent 1
Customer Lookup"]
    SUP --> A2["Agent 2
Order Processing"]
    SUP --> A3["Agent 3
Inventory Check"]
    SUP --> A4["Agent 4
Payment Processing"]

    A1 --> DB["CRM Database"]
    A2 --> OMS["Order Management"]
    A3 --> WMS["Warehouse System"]
    A4 --> PAY["Payment Gateway"]

    A1 --> SUP
    A2 --> SUP
    A3 --> SUP
    A4 --> SUP
    SUP --> RESP["Consolidated Response"]

    style USER fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style SUP fill:#e94560,stroke:#fff,color:#fff
    style A1 fill:#2c3e50,stroke:#fff,color:#fff
    style A2 fill:#2c3e50,stroke:#fff,color:#fff
    style A3 fill:#2c3e50,stroke:#fff,color:#fff
    style A4 fill:#2c3e50,stroke:#fff,color:#fff
    style RESP fill:#4CAF50,stroke:#fff,color:#fff
    style DB fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style OMS fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style WMS fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style PAY fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50

Two Orchestration Modes

import boto3

bedrock = boto3.client('bedrock-agent')

# Create Supervisor Agent
supervisor = bedrock.create_agent(
    agentName='ecommerce-supervisor',
    instruction="""You are a supervisor orchestrating an e-commerce agent team.
    Analyze customer requests and coordinate specialized agents for:
    order lookup, inventory checks, and payment processing.""",
    foundationModel='anthropic.claude-sonnet-4-20250514'
)

# Associate collaborator agents
bedrock.associate_agent_collaborator(
    agentId=supervisor['agent']['agentId'],
    collaborationInstruction='Call this agent for customer information lookup',
    collaboratorId=customer_agent_id,
    collaboratorName='CustomerLookup',
    relayConversationHistory='TO_COLLABORATOR'
)

7. Identity — Authenticating Agents with External Systems

One of the biggest challenges when deploying AI Agents to production: how do you let agents access enterprise systems (Slack, GitHub, Jira, Salesforce) securely, without hardcoding credentials?

AgentCore Identity solves this by assigning agents their own identity and integrating with corporate Identity Providers (IdP).

graph LR
    AGENT["AI Agent"] --> ACID["AgentCore Identity"]
    ACID --> IDP["Corporate IdP
Okta / Entra ID / Cognito"]
    IDP --> TOKEN["OAuth Token"]
    TOKEN --> AGENT
    AGENT --> SLACK["Slack API"]
    AGENT --> GH["GitHub API"]
    AGENT --> JIRA["Jira API"]

    style AGENT fill:#e94560,stroke:#fff,color:#fff
    style ACID fill:#2c3e50,stroke:#fff,color:#fff
    style IDP fill:#16213e,stroke:#fff,color:#fff
    style TOKEN fill:#4CAF50,stroke:#fff,color:#fff
    style SLACK fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style GH fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style JIRA fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50

8. Memory — Agents with Cross-Session Memory

AgentCore Memory automatically manages two types of memory for agents:

# Memory is automatically managed by AgentCore
# Agent can query memory via API

# Example: Agent remembers customer preferences
response = bedrock_runtime.invoke_agent(
    agentId='support-agent',
    sessionId='user-12345',
    enableTrace=True,
    memoryId='user-12345-memory',  # Long-term memory identifier
    inputText='I want to change my subscription plan'
)

# Agent automatically knows:
# - User has been on Premium plan since 2024
# - User has asked about downgrading twice before
# - User prefers Vietnamese communication

9. Gateway — Connecting Agents to Any API

AgentCore Gateway acts as a middleman between agents and external tools/APIs. Instead of each agent handling authentication, rate limiting, and error handling for every API — Gateway handles it all.

# Example OpenAPI spec for "get_order_status" tool
openapi: 3.0.0
info:
  title: Order Management API
  version: 1.0.0
paths:
  /orders/{orderId}:
    get:
      operationId: getOrderStatus
      summary: Get order status
      parameters:
        - name: orderId
          in: path
          required: true
          schema:
            type: string
      responses:
        '200':
          description: Order information
          content:
            application/json:
              schema:
                type: object
                properties:
                  status:
                    type: string
                    enum: [pending, processing, shipped, delivered]
                  estimatedDelivery:
                    type: string
                    format: date

10. Action Groups — Extending Agent Capabilities

Action Groups allow agents to perform specific actions by connecting to AWS Lambda functions or API endpoints. Each Action Group is a collection of tools the agent can invoke.

graph TB
    AGENT["Bedrock Agent"] --> AG1["Action Group: CRM"]
    AGENT --> AG2["Action Group: Inventory"]
    AGENT --> AG3["Action Group: Notifications"]
    AGENT --> KB["Knowledge Base"]

    AG1 --> L1["Lambda: searchCustomer"]
    AG1 --> L2["Lambda: updateCustomer"]
    AG2 --> L3["Lambda: checkStock"]
    AG2 --> L4["Lambda: reserveItem"]
    AG3 --> L5["Lambda: sendEmail"]
    AG3 --> L6["Lambda: sendSMS"]

    style AGENT fill:#e94560,stroke:#fff,color:#fff
    style AG1 fill:#2c3e50,stroke:#fff,color:#fff
    style AG2 fill:#2c3e50,stroke:#fff,color:#fff
    style AG3 fill:#2c3e50,stroke:#fff,color:#fff
    style KB fill:#16213e,stroke:#fff,color:#fff
    style L1 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style L2 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style L3 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style L4 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style L5 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style L6 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50

11. Observability — Monitoring Agents in Production

Agents aren't traditional APIs — output is non-deterministic, each request may go through multiple reasoning steps and tool calls. Observability in the agentic world requires tracing at the individual reasoning step level.

# Enable trace to monitor reasoning chain
response = bedrock_runtime.invoke_agent(
    agentId='my-agent',
    sessionId='session-001',
    enableTrace=True,
    inputText='Check order ORD-12345 and send an update to the customer'
)

# Trace output includes:
# 1. PreProcessingTrace - Input parsing & classification
# 2. OrchestrationTrace - Reasoning steps, tool selection
# 3. PostProcessingTrace - Response formatting
# 4. GuardrailTrace - Safety filter results
# 5. FailureTrace - Error details if any

12. Comparison with Other AI Agent Platforms

13. Production Architecture: E-commerce Support Agent

To illustrate how these building blocks combine in practice, let's examine the architecture of a Customer Support Agent system for e-commerce.

graph TB
    CUST["Customer
Web / Mobile / Chat"] --> ALB["Application Load Balancer"]
    ALB --> API["API Gateway"]
    API --> RT["AgentCore Runtime"]

    RT --> SUP["Supervisor Agent"]
    SUP --> FAQ["FAQ Agent
+ Knowledge Base"]
    SUP --> ORD["Order Agent
+ Action Groups"]
    SUP --> ESC["Escalation Agent
+ SES / SNS"]

    RT --> MEM["AgentCore Memory"]
    RT --> GR["Guardrails"]
    RT --> OBS["CloudWatch
Observability"]
    RT --> AID["AgentCore Identity
→ Okta SSO"]

    FAQ --> S3["S3: Product Docs"]
    ORD --> DDB["DynamoDB: Orders"]
    ORD --> PAY["Stripe API"]
    ESC --> AGENT_DESK["Human Agent Queue"]

    style CUST fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style RT fill:#e94560,stroke:#fff,color:#fff
    style SUP fill:#2c3e50,stroke:#fff,color:#fff
    style FAQ fill:#16213e,stroke:#fff,color:#fff
    style ORD fill:#16213e,stroke:#fff,color:#fff
    style ESC fill:#16213e,stroke:#fff,color:#fff
    style GR fill:#e94560,stroke:#fff,color:#fff
    style MEM fill:#2c3e50,stroke:#fff,color:#fff
    style OBS fill:#2c3e50,stroke:#fff,color:#fff
    style AID fill:#2c3e50,stroke:#fff,color:#fff
    style ALB fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style API fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style S3 fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style DDB fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style PAY fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50
    style AGENT_DESK fill:#f8f9fa,stroke:#2c3e50,color:#2c3e50

14. Production Best Practices

15. Cost and Pricing

Bedrock AgentCore uses a pay-per-use model, charging based on:

• LLM inference: Billed per input/output tokens, varying by model (Claude Sonnet is cheaper than Claude Opus)
• Knowledge Base: Embedding costs + vector store (OpenSearch Serverless)
• Runtime: Compute costs based on active session time
• Guardrails: Billed per text units processed
• Memory: Storage costs for long-term memory

Conclusion

Amazon Bedrock AgentCore represents the maturation of the AI Agent ecosystem — from demo stage to production-grade. With its modular architecture, you're not forced to use the entire platform — you can start with Runtime + Guardrails, then expand to Memory, Identity, and Multi-Agent as needs grow.

What makes AgentCore different isn't any single feature — Azure AI Foundry and LangGraph Cloud offer similar capabilities. The strength lies in everything being managed, serverless, and framework-agnostic — you focus on agent logic while AWS handles the production infrastructure.

References

• Amazon Bedrock AgentCore — Official Page
• AgentCore Developer Guide — AWS Documentation
• Introducing Amazon Bedrock AgentCore — AWS Blog
• Multi-Agent Collaboration — AWS Bedrock Docs
• Amazon Bedrock Guardrails — AWS
• Navigating Managed AI Agent Options on Bedrock — The Scale Factory
• AI Agents Best Practices with AgentCore — AWS ML Blog