Skip to content

⚡ Lab 3: AgentCore + MCP Server

Pattern: AgentCore Runtime + MCP
Duration: 20 minutes
Objective: Explore AWS Bedrock AgentCore with Coveo Model Context Protocol (MCP) Server for advanced tool orchestration.

Throughput Notice

If Bedrock doesn't respond immediately, wait 30-60 seconds and retry. This is normal during peak workshop usage.

🎯 Lab Goals

By the end of this lab, you will:

  • ✅ Understand AgentCore Runtime and Coveo MCP Server architecture
  • ✅ Explore AgentCore Console
  • ✅ Test multi-tool orchestration with Coveo MCP Server
  • ✅ View CloudWatch Logs for observability
  • ✅ Compare MCP with previous integration patterns
  • ✅ Identify when to use AgentCore with Coveo MCP

🏗️ Architecture Overview

In this lab, you'll explore the AgentCore with MCP pattern:

graph TB subgraph "User Interface" UI[Search UI
React Application] end subgraph "AWS Infrastructure" API[API Gateway] Lambda[AgentCore Runtime
Lambda] end subgraph "AWS Bedrock AgentCore" Runtime[Agent Runtime
Container
Nova Lite Model] Memory[AgentCore Memory
Cross-session Storage] end subgraph "MCP Server" MCP[Coveo MCP Server
Container
3 Tools] T1[search_coveo] T2[passage_retrieval] T3[answer_question] end subgraph "External Services" COVEO[Coveo Platform
Search + Passages + Answer] end UI --> API API --> Lambda Lambda --> Runtime Runtime --> MCP MCP --> T1 MCP --> T2 MCP --> T3 T1 --> COVEO T2 --> COVEO T3 --> COVEO Runtime <--> Memory style UI fill:#e1f5fe style Runtime fill:#f3e5f5 style MCP fill:#fff3e0 style COVEO fill:#e8f5e9

🔄 Pattern 3 Sequence Diagram

%%{init: {'theme':'base', 'themeVariables': { 'primaryColor':'#f3e5f5','primaryTextColor':'#000','primaryBorderColor':'#9c27b0','lineColor':'#9c27b0','secondaryColor':'#e1f5fe','tertiaryColor':'#fff3e0'}}}%% sequenceDiagram participant User participant UI as Search UI participant Runtime as AgentCore Runtime participant MCP as Coveo MCP Server participant Coveo as Coveo APIs participant Memory as AgentCore Memory User->>+UI: Question UI->>+Runtime: Invoke agent Runtime->>Runtime: Analyze question Runtime->>+MCP: Call MCP tool 1
(search_coveo) MCP->>+Coveo: Search API Coveo-->>-MCP: Search results MCP-->>-Runtime: Tool response Runtime->>+MCP: Call MCP tool 2
(passage_retrieval) MCP->>+Coveo: Passages API Coveo-->>-MCP: Relevant passages MCP-->>-Runtime: Tool response Runtime->>+MCP: Call MCP tool 3
(answer_question) MCP->>+Coveo: Answer API Coveo-->>-MCP: AI answer MCP-->>-Runtime: Tool response Runtime->>Runtime: Synthesize response Runtime->>+Memory: Store conversation Memory-->>-Runtime: Context saved Runtime-->>-UI: Comprehensive response UI-->>-User: Answer + citations Note over User,Memory: Pattern 3: AgentCore + MCP Note over Runtime,MCP: Multi-Tool Orchestration

📋 Lab Exercises

Exercise 3.1: AgentCore Architecture Overview (5 minutes)

Type: Instructor-led + Hands on

Agentcore architecture with Coveo MCP

Agentcore architecture with Coveo MCP

Key Concepts

🚀 Agent Runtime
Containerized agent orchestrator
🔌 MCP Server
Model Context Protocol server with tools
🛠️ Multiple Tools
Search, passages, and answer tools
🧠 Advanced Memory
Cross-session memory support

Why Two Runtimes in AgentCore?

AgentCore uses a separation of concerns architecture with two distinct runtime components:

🎯 Agent Runtime (Orchestrator)

  • Purpose: Decision-making and conversation management
  • Contains: Foundation model (Nova Lite), memory system, conversation logic
  • Responsibilities:
    • Analyze user questions and determine intent
    • Decide which tools to call and in what order
    • Manage conversation context and memory
    • Synthesize responses from multiple tool outputs
    • Handle session management and user interactions

🔧 Coveo MCP Server Runtime

  • Purpose: Tool execution and external service integration
  • Contains: Coveo API integrations, tool implementations
  • Responsibilities:
    • Execute specific Coveo API calls (search, passages, answer)
    • Handle API authentication and error handling
    • Format and validate tool parameters
    • Return structured responses to the Agent Runtime

Benefits of This Architecture:

Benefit Description Example
🔄 Scalability Each runtime can scale independently MCP server handles high API load while agent manages conversations
🛡️ Security Isolate sensitive operations Coveo API keys only in MCP server, not in agent
🔧 Maintainability Update tools without affecting agent logic Add new Coveo features without changing conversation flow
📊 Observability Separate logging and monitoring Track tool performance vs. conversation quality separately
🚀 Reusability MCP server can serve multiple agents Same Coveo tools used by different agent personalities

Communication Flow: 

User Question → Agent Runtime → MCP Server → Coveo APIs
                     ↓              ↓           ↓
User Response ← Agent Runtime ← MCP Server ← API Response

The Agent Runtime acts as the "brain" making decisions, while the MCP Server acts as the "hands" executing specific tasks.

What is Model Context Protocol (MCP)?

MCP is a standardized protocol for AI agents to interact with tools and services:

  • 🔌 Standardized Interface: Common protocol for tool integration
  • 🔧 Tool Discovery: Agents can discover available tools
  • 📊 Enhanced Observability: Better logging and tracing
  • 🚀 Scalability: Easy to add new tools

AgentCore vs Bedrock Agent

Aspect Bedrock Agent (Lab 2) AgentCore (Lab 3)
Deployment Managed service Containerized runtime
Protocol Proprietary MCP (standardized)
Tools 1 tool (supports multiple) Multiple tools
Orchestration Agent-driven via action groups Agent-driven via MCP
Observability Good Excellent
Flexibility Limited High

Exercise 3.2: Explore AgentCore Console (5 minutes)

Type: Hands-on

Step 1: Navigate to Bedrock AgentCore

  1. Login to AWS Console
  2. Navigate to Amazon Bedrock AgentCore
  3. Click "Agent Runtime" in the left navigation
  4. Click "workshop_CoveoAgent" to view the orchestration runtime details

AWS Console - Bedrock AgentCore Navigation

AWS Console - Bedrock AgentCore Navigation

Key Information:

Field Value Purpose
Runtime Name workshop_CoveoAgent Orchestrator Identifier
Type Orchestration Agent Protocol type
Status Ready Ready to use
Runtime ID workshop_CoveoAgent-<....> Used to Invoke Runtime

Step 2: Review Coveo MCP Server Runtime

  1. Go back to the Agent Runtime page and Find workshop_Coveo_MCPServer_Tools
  2. Click on it to view details

AgentCore Console - Coveo MCP Server Runtime Details

AgentCore Console - Coveo MCP Server Runtime Details

Key Information:

Field Value Purpose
Runtime Name workshop_Coveo_MCPServer_Tools Identifier
Type MCP Server Protocol type
Status Ready Ready to use
Runtime ID workshop_Coveo_MCPServer_Tools-<....> Used to Invoke Runtime

Step 3: Review Memory Configuration

  1. Go back to the BedrockAgentCore service page, and Find Memory, click on "Memory"
  2. Click on the memory workshop_CoveoAgent_Memory-<....> to view details

AgentCore Console - Memory Configuration

AgentCore Console - Memory Configuration

  1. Review the details of the memory configuration

AgentCore Console - Memory Details

AgentCore Console - Memory Details

Key Information:

  • Event Expiration: 7 Days
  • Long-term memory strategy Type: Summarization (built-in)
  • Namespace: /summaries/{actorId}/{sessionId}

Exercise 3.3: Test in Workshop UI (5 minutes)

Type: Hands-on

Step 1: Switch Backend Mode

  1. Return to Workshop UI
  2. Click backend selector (top right)
  3. Select "Coveo MCP Server Agent"

Workshop UI - Coveo MCP Backend Selected

Workshop UI - Coveo MCP Backend Selected

Step 2: Run Sample Queries

Test these queries and observe the responses:

Query 1: Basic Question 

managing retirement distribution planning

Expected: Single Shot Answer response with answer tool selection ( we can verify the tool selection in AWS Cloud watch logs)

Query 2: Multi-part Question 

comparing college financing options FAFSA and 529
Expected: Agent Orchestrator uses the answer tool, fails to get a single answer and falls back to Passage tool to retrive the relevant passage and sumamrizes and present the answer.

Query 3: Procedural Question 

capital gain and dividend reinvestment strategy
Expected: Agent Orchestrator uses the answer tool for this query

Exercise 3.4: View CloudWatch Logs (5 minutes)

Type: Hands On

The instructor will demonstrate how to view CloudWatch Logs for observability.

Step 1: Navigate to CloudWatch

1. Open CloudWatch in a new tab

2. Click "Logs" → "Log groups"

3. Find log groups: expand the log to look at the Agent runtime tool selection and reasoning and how it responded back

  • /aws/bedrock/modelinvocations/StackSet-workshop-layer3-ai-services-<some random number>

4. Select the log stream aws/bedrock/modelinvocations: expand the latest log based on the timestamp to view the details of the model invocation

CloudWatch - Log Groups List

CloudWatch - Log Groups List

Step 2: View Agent Runtime Logs

1. Go back to Log groups

2. Click on /aws/bedrock-agentcore/runtimes/workshop_CoveoAgent-<randomnumber>-DEFAULT

3. Click on the latest log stream

4. Observe agent logs:

  • Model Used
  • MCP Runtime ARN
  • Memory usage by the Coveo Agent
  • Unique memory_id

CloudWatch - Agent Runtime Logs

CloudWatch - Agent Runtime Logs

🔍 Key Observations

AgentCore + MCP Strengths

🔧 Multi-Tool Orchestration

Agent can use multiple tools in sequence for comprehensive answers

📊 Enhanced Observability

Detailed logs show tool selection and execution

🔌 Standardized Protocol

MCP provides consistent tool interface

🚀 Scalable Architecture

Easy to add new tools and capabilities

Comparison with Previous Labs

Aspect Coveo Direct (Lab 1) Bedrock Agent (Lab 2) AgentCore MCP (Lab 3)
Tools Available N/A 1 tool (in this workshop) 3 tools
Tool Selection Manual Automatic Intelligent
Multi-tool Use No Yes (capable) Yes
Observability Basic Good Excellent
Complexity Low Medium High
Flexibility Limited Moderate High
Response Quality Good Better Best

💡 When to Use AgentCore + MCP

Ideal Use Cases

🔬 Complex Research

Multi-step research requiring different types of information

Example: "Compare investment options, analyze risks, and recommend a strategy"

💼 Consultation Services

Expert advice requiring multiple data sources and analysis

Example: Financial planning with tax implications and investment strategies

📊 Data Analysis

Combining search, retrieval, and analysis tools

Example: Market research with trend analysis and recommendations

🎓 Advanced Education

Teaching complex topics with multiple perspectives

Example: Explaining financial concepts with examples and comparisons

✅ Validation Checklist

Before proceeding to Lab 4, verify:

  • Viewed AgentCore Runtimes in AWS Console
  • Reviewed MCP Server Runtime configuration
  • Reviewed Agent Runtime configuration
  • Tested in Workshop UI with Coveo MCP backend
  • Viewed CloudWatch Logs for both runtimes
  • Observed tool orchestration in logs
  • Understand multi-tool capabilities
  • Understand when to use AgentCore + MCP

🎉 Lab 3 Complete!

You've successfully:

  • ✅ Explored AgentCore Runtime with Coveo MCP Server architecture
  • ✅ Reviewed runtime configurations in AWS Console
  • ✅ Viewed CloudWatch Logs for observability
  • ✅ Tested multi-tool orchestration
  • ✅ Understood when AgentCore + MCP is the optimal choice

Key Takeaways

  1. AgentCore provides containerized agent runtime with advanced capabilities
  2. Coveo MCP server can be used in Agentcore to call Coveo APIs as tools
  3. MCP standardizes tool integration for better interoperability
  4. Multiple tools enable comprehensive responses through orchestration
  5. Enhanced observability through detailed CloudWatch Logs
  6. Scalable architecture makes it easy to add new capabilities

🚀 Next Steps

Ready to see all three backends in action with multi-turn conversations?

In Lab 4, you'll test multi-turn conversations and compare all three integration patterns!