aws-samples/bedrock-agentcore

Amazon Bedrock AgentCore Guide

Enterprise-grade services for deploying and operating AI agents at scale. AgentCore eliminates undifferentiated infrastructure work — identity, memory, tool management, observability — so you can focus on agent logic.

AgentCore works with any framework (Strands, LangGraph, CrewAI, LlamaIndex) and any foundation model in or outside Amazon Bedrock.

Why AgentCore?

• Faster time to value — Fully-managed services eliminate infrastructure complexity
• Flexibility and choice — Any framework, any model, full control over agent operations
• Security and trust — Enterprise-grade security, complete session isolation, comprehensive controls

AgentCore Services

1. Runtime

Securely deploy and run agents at scale. Runtime provides:

• Containerized agent execution with automatic scaling
• Session isolation between concurrent users
• Health monitoring and automatic restarts
• Support for any Python-based agent framework

Football context: Runtime hosts your football agent so it can respond to game ticks in real-time. Each deployed agent gets a unique Runtime ARN used for match registration.

Configuration file: .bedrock_agentcore.yaml in your agent project root defines runtime settings.

2. Memory

Context-aware agents with managed memory infrastructure:

• Short-term memory — Multi-turn conversation context within a session
• Long-term memory — Persistent knowledge shared across agents and sessions
• Industry-leading accuracy for memory retrieval

Football context: Your agent can remember opponent tendencies from previous matches (long-term) and track patterns within the current match (short-term). Memory can be shared across your team — when one defender learns an opponent's dribbling pattern, all defenders gain that knowledge.

3. Gateway

Secure tool discovery and management for agents:

• Transform APIs, Lambda functions, and existing services into agent-compatible tools
• Built-in semantic search to find the right tool for the context
• Support for thousands of tools while minimizing prompt size and latency
• Eliminates weeks of custom integration code

Football context: Instead of hardcoding every tool into every player agent, Gateway lets agents discover the right tool for the moment — a midfielder under pressure finds the "quick pass" tool, while a striker in the box discovers the "power shot" tool.

4. Identity

Secure agent identity and access management:

• Compatible with existing identity providers (no user migration needed)
• Secure token vault to minimize consent fatigue
• Just-enough access and secure permission delegation
• Agents can securely access AWS resources and third-party services

Football context: Each player agent gets a unique identity with role-specific permissions. Your striker accesses shooting tools, your goalkeeper controls the penalty area, and your midfielder accesses passing tools — each agent only accesses actions relevant to their role.

5. Observability

Trace, debug, and monitor agent performance in production:

• OpenTelemetry-compatible telemetry
• Detailed visualizations of each step in the agent workflow
• Unified operational dashboards
• Quality standards monitoring at scale

Football context: Trace your agent's decision-making step by step — which tools it called, what data it received, where logic failed. Monitor all 5 player agents simultaneously, identify performance bottlenecks, and optimize responsiveness.

6. Code Interpreter

Secure code execution in isolated sandbox environments:

• Run Python code in sandboxed containers
• Advanced configuration support
• Seamless integration with popular frameworks
• Enterprise security requirements met

Football context: Your agent can run physics calculations in real-time — given positions of players, goalkeeper, and defenders, calculate the optimal shooting angle and power dynamically instead of pre-programming every scenario.

7. Browser

Cloud-based browser runtime for AI agents:

• Fast, secure, scalable browser interactions
• Enterprise-grade security and comprehensive observability
• Automatic scaling without infrastructure management

Football context: Agents can browse tactical analysis sites, watch video clips of professional players, and extract positioning patterns to bring real-world football intelligence into your simulation.

Deploying Your Football Agent

Prerequisites

• Agent code with .bedrock_agentcore.yaml config file
• AWS credentials configured (from Workshop Studio)
• AgentCore CLI installed

Step 1: Test Locally

# Test without LLM (rule-based fallback)
python test_local.py

# Test with LLM (Bedrock model)
python test_local.py --llm

Step 2: Install AgentCore CLI

Install from the Bedrock AgentCore Starter Toolkit:

https://github.com/aws/bedrock-agentcore-starter-toolkit

Step 3: Deploy

agentcore deploy

This packages your agent code, uploads it to AgentCore Runtime, and starts the service. On success you receive a Runtime ARN — a unique Amazon Resource Name identifying your deployed agent.

Step 4: Get Your Runtime ARN

Find your Runtime ARN in the Bedrock AgentCore console:

https://us-east-1.console.aws.amazon.com/bedrock-agentcore/agents?region=us-east-1

The Runtime ARN looks like:

arn:aws:bedrock-agentcore:<region>:<account-id>:runtime/<runtime-id>

Step 5: Register for Matches

Use the Runtime ARN to register your agent for football matches. Options:

• Single ARN — Use one Runtime ARN for all 5 players (same agent logic for every position)
• Multiple ARNs — Deploy separate agents per player for position-specific strategies

Step 6: View Logs

Monitor your deployed agent with CloudWatch Logs:

aws logs tail /aws/bedrock-agentcore/runtimes/<runtime-id>-DEFAULT \
  --log-stream-name-prefix "$(date +%Y/%m/%d)/[runtime-logs" \
  --follow

This streams real-time logs showing:

• Agent invocations and responses
• Tool calls and results
• Errors and exceptions
• Latency metrics

Step 7: Iterate

After watching your agent play:

1. Analyze logs for decision-making issues
2. Update agent code locally
3. Test with python test_local.py --llm
4. Redeploy with agentcore deploy
5. Re-register if you get a new Runtime ARN

AgentCore Configuration

.bedrock_agentcore.yaml

This file in your agent project root configures the Runtime deployment:

# Example configuration
runtime:
  name: my-football-agent
  entry_point: src/main.py
  requirements: requirements.txt

Logging and Observability Configuration

Enable detailed logging in your agent code to leverage AgentCore Observability:

For Strands agents:

import logging
logging.getLogger("strands").setLevel(logging.DEBUG)
logging.basicConfig(
    format="%(levelname)s | %(name)s | %(message)s",
    handlers=[logging.StreamHandler()]
)

For LangGraph agents:

import logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Log key decision points
logger.info(f"Game state received: {state_summary}")
logger.info(f"Agent action: {action}")

Metrics via Strands SDK:

result = agent("Analyze game state...")
summary = result.metrics.get_summary()
print(f"Tokens used: {summary['accumulated_usage']['totalTokens']}")
print(f"Latency: {summary['accumulated_metrics']['latencyMs']}ms")
print(f"Tool calls: {summary['tool_usage']}")

Runtime ARN Usage for Match Registration

After deployment, the Runtime ARN is your agent's identity in the competition system:

1. Copy the ARN from the AgentCore console
2. Register via the workshop match registration page
3. Assign positions — map the ARN to player slots (1-5)
4. Monitor — watch matches and check logs for your agent's decisions

You can update your agent code and redeploy without changing the ARN (the runtime endpoint stays the same).

Troubleshooting

| Issue | Solution | |-------|----------| | agentcore deploy fails | Check AWS credentials: aws sts get-caller-identity | | Agent times out during match | Ensure response time < 500ms; use a faster model (Nova Micro) | | No logs appearing | Verify the runtime ID in the log group path; check region | | Runtime ARN not found | Check the AgentCore console; deployment may still be in progress | | Model throttling errors | Add retry strategy in agent code; reduce token usage |

Resources

• Amazon Bedrock AgentCore Developer Guide
• AgentCore Starter Toolkit
• AgentCore FAQs
• AgentCore Runtime Deep Dive (YouTube)
• AgentCore Gateway Deep Dive (YouTube)
• AgentCore Memory Deep Dive (YouTube)
• AgentCore Observability (YouTube)
• AgentCore Identity (YouTube)