(AWS re:invent 2025) AgentCore Overview

AWS Bedrock AgentCore is the "Infrastructure-as-Code" layer for AI agents. It wraps your raw python code with enterprise-grade security and scaling. This guide summarizes the key components: Runtime, Identity, Gateway, and Observability. ## 1. Agent Runtime (Serverless Hosting) You can turn any Python script into a scalable microservice with just 4 lines of configuration. ```python from bedrock_agentcore_starter_toolkit import Runtime # Auto-configure IAM, ECR, and API Gateway response = agentcore_runtime.configure( entrypoint="app.py", auto_create_execution_role=True, auto_create_ecr=True ) ``` ## 2. Identity & Security Security is handled via decorators. Never hardcode API keys. ### Outbound Security (API Keys) Use the credential provider to inject keys only when needed. ```python @requires_api_key(provider_name="openai-apikey-provider") async def call_openai(*, api_key: str): # The key is injected safely at runtime client = OpenAI(api_key=api_key) ``` ## 3. AgentCore Gateway (The "Librarian") If you have 300 tools, you can't feed them all to the LLM. The Gateway uses **Semantic Search** to find the 5 most relevant tools for the current query and presents only those to the model, reducing hallucinations and token costs. ## 4. Memory Hooks To ensure compliance, use **Memory Hooks** to automatically capture audit logs or user preferences. ```python class AuditHook(HookProvider): def on_agent_initialized(self, event): # Automatically load recent history self.memory_client.get_last_k_turns(5) ``` ## 5. Observability (OpenTelemetry) AgentCore has built-in OTEL support. You just need to instrument your Docker container. **Dockerfile:** ```dockerfile # Enable auto-instrumentation CMD ["opentelemetry-instrument", "python", "-m", "my_agent"] ``` **Result:** You get full X-Ray traces in the AWS Console, showing exactly how long each tool took to execute and where any errors occurred. **Conclusion:** AgentCore solves the "Day 2" problems of AI: deployment, security, and monitoring. It allows developers to focus on the prompt, not the plumbing.

1
2
3
4
5
6
7
8

from bedrock_agentcore_starter_toolkit import Runtime

# Auto-configure IAM, ECR, and API Gateway
response = agentcore_runtime.configure(
    entrypoint="app.py",
    auto_create_execution_role=True,
    auto_create_ecr=True
)

1
2
3
4

@requires_api_key(provider_name="openai-apikey-provider")
async def call_openai(*, api_key: str):
    # The key is injected safely at runtime
    client = OpenAI(api_key=api_key)

1
2
3
4

class AuditHook(HookProvider):
    def on_agent_initialized(self, event):
        # Automatically load recent history
        self.memory_client.get_last_k_turns(5)

1
2

# Enable auto-instrumentation
CMD ["opentelemetry-instrument", "python", "-m", "my_agent"]