Vertex Macro｜Financial Cloud Cloud · AWS Amarathon 2025
Vertex Macro｜Amarathon 2025 Recap 36：Making Cloud Cost Analysis Smarter: Building FinOps Intelligent Agents with Strands and AgentCore
The year 2025 is known as the "Year One of AI Agents," and it's just two months away.
● Users have already developed their own AI Agents.
● Programming agents are a type of AI agent.
● AI agents are various "smart-looking applications" that utilize AI.
● Examples include: meeting minutes agents, interview preparation agents, and programming agents.
Characteristics of "AI Agents" in the LLM era
● Role profiling: can define roles or personalities, achieving personalized behavior and responses.
● Planning and reflection: to achieve goals, agents can formulate plans and make adjustments based on execution results.
● Long-term memory: can retain long-term interaction information or experiences like humans.
● Tool execution: can not only generate text but also call various external tools or APIs to perform operations.
● AI Agents can help solve practical problems, such as cost analysis agents.
● How to use Amazon Web Services to develop cost analysis agents.
How to build AI Agents
● Bedrock Agents
● AgentCore: deploy self-developed agents in a serverless manner, providing authentication, tools, observability, and other functions.
● Strands Agents: a framework for Python, requiring only a minimum of 3 lines of ultra-concise code to implement an AI agent.
● Building FinOps agents with Strands Agents.
Open-source AI Agent framework—Strands Agents
● Create AI agents with just 3 lines of Python code
● Advantages: simple, lightweight, good development experience
● Applied to Amazon Web Services, such as Amazon Q Developer
● Strands core concept: combining "models" and "tools"
● With the improvement of LLM capabilities, building AI agents only requires specifying models and tools
● Agent creation: set LLM and system prompts, and call by providing prompts
● Equip agents with tool capabilities: Strands provides built-in tools, such as calculation and file operations
● You can write your own tools by adding @tool to Python functions
● Obtain tools provided by MCP servers, compatible with local and remote MCP servers
● Build multiple agents: Agent as Tools, Swarm, Graphs, Workflows
● Agent as Tools example: supervisor-subordinate model, where the supervisor agent assigns sub-tasks, calls sub-agents to execute, and summarizes results
● Popular trend after MCP: A2A (Agent to Agent) support
● A2A on Strands construction example: there are specialized classes that can call remote agents using Tool Use
● Agent deployment: quickly deploy through Bedrock AgentCore
Troubles of AI Agent deployment
● Complicated deployment process
● Authentication and authorization issues
● Maintenance and monitoring challenges
● Running cost issues
● Whether streaming output is supported
What is Bedrock AgentCore?
● It is a "convenient component set" dedicated to AI agents
● Includes functions such as Runtime (serverless infrastructure), Memory (memory management), Gateway (tool integration), Identity (authentication/authorization), and Observability (maintenance and monitoring)
● Can be used with any preferred agent framework, selecting functions as needed, and easily integrated through APIs
The core of the entire system is the runtime
● Regardless of the framework used to develop AI agents, they can be easily deployed in a serverless environment
● Similar to containerized Lambda specifically prepared for AI agents
● With the help of a dedicated CLI toolkit, deployment operations can be easily completed
● The backend part is completed, and the next step is to consider the implementation plan for the frontend
Optional technologies for the frontend page
● Beginner-friendly
● Production-level development suite
● You can easily write a beautiful interface with Python
● Simply associate code repositories such as Next.js or React to achieve automatic deployment
● For backend engineers who are not proficient in JavaScript, this is a good choice
● The Gen2 version has achieved significant evolution and performance improvement.
Key Takeaway
● Cloud cost analysis is crucial for both enterprises and individuals. In the Year One of AI Agents, you can quickly build AI agents that help analyze cloud costs using Amazon Web Services' technology stack.
● Strands Agents is a framework for flexibly building multi-agent systems, requiring only 3 lines of Python code to set up AI agents, featuring extreme simplicity and high scalability.
● Strands Agents support MCP and A2A protocols, enabling collaboration and tool sharing among agents.
● Strands Agents can seamlessly integrate with Bedrock AgentCore to achieve production-level deployment.
● Using AgentCore can simplify the deployment and maintenance process. The Runtime component provides a serverless runtime environment with automatic scaling.
● Components such as Memory, Identity, Gateway, and Observability provide integrated capabilities for memory, authentication, tool integration, and monitoring.
● Automatic packaging and deployment through CLI can quickly enter the production environment.
Amazon Bedrock AgentCore architecture
Core Services
● AgentCore Runtime: A secure, serverless execution environment that hosts your AI agent or tool code. It offers complete session isolation for security and supports long-running asynchronous tasks up to 8 hours.
● Framework: Supports popular open-source agent frameworks (e.g., LangGraph, CrewAI) and any foundation model.
● Agent Instructions: Defines the behavior and capabilities of the agent.
● Agent Local Tools: Tools that are local to the specific agent for performing tasks.
● Agent Context: Manages the ephemeral, session-specific state within a conversation.
● AgentCore Gateway: Provides a secure way for agents to discover and connect with tools and resources. It can transform existing APIs (like Lambda functions or OpenAPI specs) into agent-compatible tools, minimizing custom integration work.
● AgentCore Memory: Enables agents to have context-aware conversations by managing both short-term and long-term memory. It stores conversational context and extracts persistent knowledge like user preferences across sessions.
● AgentCore Identity: Offers secure, scalable identity and access management for agents. It handles authentication and authorization, allowing agents to securely access AWS resources and third-party services on behalf of users.
● CloudWatch GenAI Observability (AgentCore Observability): Provides comprehensive monitoring, tracing, and debugging capabilities for agent performance in production. It offers deep operational insights into the agent's workflow, powered by Amazon CloudWatch and OpenTelemetry compatible telemetry.
Built-in Tools
● AgentCore Code Interpreter: Allows agents to write and execute code securely in isolated sandbox environments for complex tasks like data analysis or calculations.
● AgentCore Browser: Provides a fast, secure, cloud-based browser runtime for agents to interact with and extract information from websites at scale.
External Interactions
● App & Models: The agent system interacts with user applications and various foundation models (FMs) from Amazon Bedrock or other providers to perform its tasks.
Amazon Bedrock AgentCore starter toolkit
● Code: The process starts with the source code for the AI agent.
● Build: The "agentcore launch" command within the toolkit automatically triggers an AWS CodeBuild project.
● Container: CodeBuild compiles the agent code into a container image, optimized for the environment (e.g., ARM64 architecture).
● ECR (Elastic Container Registry): The built container image is "Pushed" to an Amazon ECR repository, which serves as persistent storage for the image.
● AgentCore Runtime: The image is then deployed to the secure, serverless Amazon Bedrock AgentCore Runtime, the execution environment for the AI agent.
● X-Ray: The system integrates with AWS X-Ray for observability, providing tracing and debugging capabilities for the agent's performance in production.
● Automation: The bottom text indicates that the agentcore-starter-toolkit automatically handles these configuration and deployment steps .