Agentic AI has emerged as one of the most exciting topics in today's generative AI landscape. But as forward-thinking developers, we want to go beyond just understanding concepts and terminologies. We want to see real innovations in action and learn how to build our own AI agent applications from the ground up.
That's exactly what Haowen Huang (Senior Developer Advocate, AWS) and Jacky Wu (Senior Solutions Architect, AWS) delivered in their DEV327 session at AWS re:Invent 2025. Both based in Hong Kong, they brought practical, hands-on demonstrations showing how to build agentic AI applications using Amazon Nova Act and the open-source Strands Agents SDK.
Watch the Full Session:
The Evolution of Agentic AI
Haowen opened by examining the evolution of generative AI, which has followed a clear progression. On one end, we have low-agency, rule-based systems requiring high human oversight. On the other end, we see high-agency systems that can operate independently and even make strategic decisions autonomously.
"Although we are still in the very early stage of agentic AI, we have a vision of a future world powered by AI agents," Haowen explained, referencing Sequoia Capital's keynote at AI Ascent 2025.
Sequoia Capital predicts that future AI systems will evolve beyond current tool-like attributes into autonomously operating intelligent agents with capabilities for reasoning, planning, collaboration, and high autonomous operation. Their vision for the 2030s includes an "agent economy" operating like a global neural network composed of numerous interconnected agent operations.
The implications are striking. Sequoia Capital predicts the emergence of the first one-person unicorn - a company created and operated by a single individual with a valuation of $1 billion. This shift in labor structure gives rise to entirely new organization models where a business can be run by a single person while performing all the functional capabilities at the scale of a complete startup.
Building the Foundation: Communication Protocols
Just as TCP/IP laid the foundation for network interconnections during the development of the internet, AI agents require interoperability protocols. Emerging standards like Model Context Protocol (MCP) and Agent-to-Agent (A2A) are addressing these challenges. AWS is actively participating in the communities developing both MCP and A2A protocols, contributing Amazon's decades of experience in distributed systems to enhance these future standards for agent communication.
Real-World Demo: Hong Kong Weather Agent with Amazon Nova Act
Haowen demonstrated a practical AI agent that automatically retrieves local weather information from the Hong Kong Observatory website. The demo showcased how Amazon Nova Act can function like a human web-scraping engineer, autonomously locating information and saving verification screenshots.
The implementation is remarkably straightforward. The Nova Act forecast function guides Amazon Nova Act on which specific web pages to capture and what actions to perform. Using natural language, developers can assign tasks like "read and extract the complete nine-day weather forecast information visible on this page" along with the Hong Kong Observatory website URL as the starting page.
The code demonstrates a typical MCP call using the MCP decorator and the process of running Nova Act in a separate thread. By simply changing the URL, this weather agent can easily be adapted for use in other countries.
Educational AI: Mathematics Video Creation with Strands Agents SDK
The second demonstration showcased the creation of mathematical animations using Manim, a Python library designed specifically for high-quality mathematical visualization. The Strands Agents SDK processed a user prompt, wrote a precise Manim script, and generated a polished animation demonstrating the powerful capabilities of combining Strands Agents SDK with Python libraries.
The implementation required minimal code:
Import Essential Components - The code imports the agent class from the Strands module and MCP client class for establishing connections with the Manim MCP server.
Establish MCP Connection - A connection to the Manim MCP server is set up using input and output transport mechanisms.
Prepare Tools - The system chooses available tools from the MCP server and prepares the agent with these tools for upcoming tasks.
Process Natural Language Prompt - The agent processes a natural language request like "create a Manim animation that visualizes a cubic function graph from X = -3 to X = 3."
During the demo, when the agent detected compatibility issues in the local environment, it intelligently created a simplified version to complete the task, demonstrating its adaptive problem-solving capabilities. The entire educational video creation process took just 30 minutes to develop.
What is Strands Agents SDK?
Strands Agents SDK is an open-source framework created by AWS that enables developers to build AI agents with minimal code. The framework simplifies development by eliminating complex agent orchestrations and leverages state-of-the-art large language models to handle planning, chain of thought, tool calling, and more.
Building a truly functional AI agent requires substantial work, with potentially 80-90% of the effort being unrelated to unique business logic. Strands Agents SDK addresses this challenge by providing a comprehensive framework that handles the foundational infrastructure.
Multi-Agent Collaboration Patterns
Jacky Wu took the session deeper into multi-agent systems, explaining why multi-agent architectures matter:
Specialization - Specialized agents working in specific domains increase precision and deliver superior outcomes when combined.
Scalability - Agents can be added or removed without impacting the whole system during the workflow.
Maintainability - Smaller agents are easier to control, monitor, debug, and test (debugging 100 lines of code versus 10,000 lines makes a significant difference).
Cost Efficiency - Different agents can use different models matched to task complexity, balancing speed, performance, and cost.
"Multi-agent systems are like a well-coordinated team," Jacky explained. "A monolithic agent is like a trader managing a portfolio all by himself. But multi-agent is like a team where you can have analysts doing market research, quants doing modeling, risk managers managing portfolio risk, and IT maintaining infrastructure. When they work together, they achieve superior performance."
Four Multi-Agent Collaboration Patterns
Strands Agents SDK supports four key multi-agent collaboration patterns:
1. Agent as Tools
An orchestrator agent receives user requirements, understands intent, and calls specialized agents to perform specific tasks. The output of one agent becomes the input for the next agent in the chain.
Jacky demonstrated this pattern with a quantitative hedge fund use case where a quant research agent orchestrates market data tools, backtest tools, strategy generation agents, and results summary agents. The demo showed how researchers can verify trading ideas quickly using historical market data, with the system providing profit/loss metrics, Sharpe ratios, buy/sell conditions, and improvement suggestions.
The implementation leveraged Amazon SageMaker AgentCore Gateway to build tools from Lambda functions and AgentCore Runtime to host agents securely. Different language models were used for different sub-agents - Claude Sonnet for coding capacity in the strategy generation agent and Amazon Nova for summarization tasks.
2. Swarm
In the Swarm pattern, there's no single orchestrator. Agents communicate with each other and collaborate dynamically to solve problems. The Swarm maintains a shared context that all agents can access, including information about who is in the system, what other agents input and output, and historical interactions.
Jacky demonstrated a blog creation example with research, creative, critical, and summarize agents. The interesting behavior emerged when the summarize agent determined it needed more details from the creative agent to create an effective social media summary. The agent reasoned: "I see the general structure of the blog post from the creative agent's descriptions, but I need more details to create an effective social media summary." It then handed control back to the creative agent to fill in the content before completing the task.
3. Graph
In the graph pattern, agent relationships are defined as a graph network where each agent is a node and edges define relationships and information flow. The key advantage is predictability and control - you can specify exactly which agents can communicate with which other agents.
Jacky showcased an agentic fund manager developed by AWS colleague Kang Wan Lee, demonstrating how Strands Agents SDK works with other popular agentic frameworks like LangGraph. The implementation defined nodes for financial analysis, portfolio construction, and risk management, with edges controlling the information flow between them.
4. Workflow
The workflow pattern provides a well-defined sequence of tasks with clear dependencies. Tasks can specify which other tasks they depend on, and priorities determine execution order when multiple sub-agents can run in parallel.
This pattern is perfect for complex, repeatable processes like automated data pipelines with extraction, transformation, and loading steps.
Practical Use Cases Across Industries
The session concluded with practical use cases for each pattern:
Agent as Tools - Customer service chatbots where an orchestrator calls technical account managers, billing agents, and summarize agents to provide comprehensive customer responses.
Swarm - Multidisciplinary incident response where agents dynamically communicate to diagnose and solve problems across networking, database, and application layers.
Graph - Data validation with error paths where you clearly define which agents handle error-free data versus data requiring correction.
Workflow - Automated data pipelines with extraction, ETL, transformation, and loading into data warehouses.
Key Takeaways
Start with Clear Use Cases - Both demos showed practical, real-world applications rather than theoretical concepts. The weather agent and mathematics video creator demonstrate how quickly you can build functional AI agents.
Leverage Open Source - Strands Agents SDK is an open-source framework created by AWS, providing developers with source code level control while eliminating 80-90% of the boilerplate work.
Choose the Right Pattern - Understanding the four multi-agent collaboration patterns helps you architect solutions that match your specific requirements for control, flexibility, and scalability.
Match Models to Tasks - Multi-agent systems let you use different language models for different sub-agents, optimizing for speed, performance, or cost based on task complexity.
Embrace the Stochastic Mindset - Success in the new era of agentic AI requires mental preparation beyond just technology. Consider organizational changes and new mental frameworks for working with AI systems.
Protocols Matter - Just as TCP/IP enabled the internet, MCP and A2A protocols are laying the foundation for the future agent economy.
As Haowen emphasized, "Whether you embrace it or not, the era of agentic AI has arrived." The tools and frameworks showcased in this session - Amazon Nova Act, Strands Agents SDK, AgentCore Gateway, and AgentCore Runtime - provide developers with everything needed to start building production-ready agentic AI applications today.
About This Series
This post is part of DEV Track Spotlight, a series highlighting the incredible sessions from the AWS re:Invent 2025 Developer Community (DEV) track.
The DEV track featured 60 unique sessions delivered by 93 speakers from the AWS Community - including AWS Heroes, AWS Community Builders, and AWS User Group Leaders - alongside speakers from AWS and Amazon. These sessions covered cutting-edge topics including:
- 🤖 GenAI & Agentic AI - Multi-agent systems, Strands Agents SDK, Amazon Bedrock
- 🛠️ Developer Tools - Kiro, Kiro CLI, Amazon Q Developer, AI-driven development
- 🔒 Security - AI agent security, container security, automated remediation
- 🏗️ Infrastructure - Serverless, containers, edge computing, observability
- ⚡ Modernization - Legacy app transformation, CI/CD, feature flags
- 📊 Data - Amazon Aurora DSQL, real-time processing, vector databases
Each post in this series dives deep into one session, sharing key insights, practical takeaways, and links to the full recordings. Whether you attended re:Invent or are catching up remotely, these sessions represent the best of our developer community sharing real code, real demos, and real learnings.
Follow along as we spotlight these amazing sessions and celebrate the speakers who made the DEV track what it was!
Top comments (0)