Vertex Macro｜Financial Cloud Cloud · AWS Amarathon 2025
Vertex Macro｜Amarathon 2025 Recap 01：A Developer’s Roadmap to Architecting for Agents
Developer’s roadmap
● future-architect
● Content generation
● Learning RAG
● Amplifying development skills
● Building AI Agent
● Integrating MCP
● Agentic communication
● It’s growing at a fast pace...
Agentic Development
● Developing WITH agent
● Code generation
● AI-DLC (AI-Driven Development Lifecycle, a modern software development methodology that positions artificial intelligence as a central collaborator)
● Agentic IDE
● Vibe coding
● Spec-driven
Developing FOR agent
● Scaling
● Deployment
Developing Agent
● Single-agent vs Multi-Agent
● MCP
● Guardrails
● Observability
● Best practices
● API
Agentic AI Systems
● Built Using
● Agent
● Autonomous Decision Making
● Iterative Problem Solving
● Goal-Oriented Behavior
Agentic Patterns
● Pattern Types
● Reflection
● Tool Use
● Planning
● Multi-Agent
Content Generation
● Choose your models
● Prompt engineering
● Parameters
● How to interact with Bedrock API
RAG with Amazon Bedrock Knowledge Base
● Use RetrieveAndGenerate API
● Setup your RAG
● Choose suitable model
KIRO
● The AI IDE for prototype to production
● Spec based Development
● Steering ( the process of guiding or controlling an AI system's behavior, responses, and development in a desired direction. This can involve fine-tuning models for specific tasks)
● Automate with Agent Hooks (automated triggers within a development environment (IDE) that execute predefined AI agent actions in response to specific events, such as saving or creating a file. They are designed to automate repetitive tasks, ensure consistency, and streamline the development workflow)
● Checkpointning
● Property-based testing - PBT (software testing methodology that focuses on verifying general properties or invariants of a system under test, rather than checking specific examples with predefined inputs and expected outputs. PBT utilizes a generative engine to automatically create diverse and randomized inputs to thoroughly explore the input space.)
AI through agent is changing development
● Agentic Development
● Developing WITH agent
● Developing FOR agent
● Developing Agent
Multi-agent patterns landscape
● Graph
● Workflow
● Swarm (allows specialized agents to work together by handing off control to one another, creating more complex and robust workflows than a single agent could achieve alone.)
Multi-agent — Swarm
● Dynamic handoffs between specialized agents
● Emergent paths - agents decide who to hand off to next
● Shared context and working memory across all agents
● Autonomous collaboration with minimal orchestration
● Use Case: Development projects, research projects
“What’s the future look like with GenAI?”
● The future of microservices isn't just about better APIs - It's about intelligent services that communicate through AI agents
● GenAI can write your code and run your workflows, but it can't replace your understanding of why the system needs to exist in the first place.
Agentic communication proof of concept
● This architecture diagram illustrates an AWS-based system for agentic AI applications using Amazon Bedrock AgentCore Runtime and Strands Agents SDK. The system utilizes the Model Context Protocol (MCP) to integrate with various microservices implemented via AWS Lambda and Amazon DynamoDB.
● AgentCore Runtime: The core execution platform for running AI agents with enterprise-grade features such as session isolation (using microVMs), scalability, and observability.
● Strands Agents: An open-source, code-first Python SDK for building the agent's logic, including handling state, tool orchestration, and multi-step reasoning.
● AgentCore Gateway: Provides secure ingress connectivity and a unified interface for agents to access tools, including existing MCP servers, REST APIs, and Lambda functions.
● Model Context Protocol (MCP): An open standard and client-server architecture that enables AI models to communicate with external data sources and tools.
● MCP Servers: Lightweight programs that expose specific capabilities (tools) to the AI agent.
● Microservices: The system is composed of several serverless microservices), each implemented using AWS Lambda functions and backed by Amazon DynamoDB for data persistence.
● Workflow: The Strands Agents running in AgentCore Runtime can make decisions and use tools by communicating through the AgentCore Gateway to invoke the various MCP servers, which in turn trigger the relevant microservices.