Building a Multi-Agent Research Team with Strands Agents

Today I’m building a multi-agent research team to retrieve recent news and produce a business report. I’ll use Strands Agents, an open-source framework from AWS that supports different model providers and tools, including MCP. AWS sponsored this video. The framework’s code is free to use, but model calls and cloud services may still cost money.

I start from the Strands GitHub repository and open a clean project in Cursor. The project contains a requirements.txt file for Python dependencies, an .env file for configuration and a strands_demo.py file. I’ll run a basic example first, then build the more involved multi-agent application.

The introductory example imports Agent and a calculator tool, creates an agent with that tool, asks for the square root of 1764 and prints the result. Strands also provides other tools, and developers can write their own Python functions or connect supported MCP servers. The exact packages and interfaces should be checked against the version being installed.

Running the initial example without credentials produces an error. In the configuration shown, Strands uses an Amazon Bedrock model by default, so I demonstrate granting Bedrock permissions in AWS IAM and creating credentials. Other model providers can be configured instead; access and model availability depend on account, region and provider.

The video puts IAM access keys in an .env file and then reruns the script. AWS recommends temporary credentials and least-privilege access where possible; secrets should be kept out of source control. The configured calculator example returns 42 for the square root of 1764.

To choose a different Bedrock model, I look up its model ID in the AWS console. I also demonstrate configuring an OpenAI model and passing it to the agent, using GPT-3.5 Turbo as the historical example. Console menus, model availability and provider integration details can change, so consult current documentation when reproducing it.

Next I create custom Python tools using the @tool decorator. The function’s docstring describes its purpose to the agent; for example, an add_numbers tool takes two numbers and returns their sum. When I run it with five and three, the tool produces eight.

I extend the example with a math agent and a text agent. The former can add and multiply, while the latter counts characters and words. The demonstration adds 15 and 25, counts the four words in ‘AWS Strands is powerful,’ and passes that count back to the math agent to multiply by two, yielding eight.

The larger project creates business-intelligence reports by coordinating specialists. I outline a content agent for news, a social-media analyst, a research analyst, strategic and sentiment specialists, a recommendations agent and an executive synthesizer. The social-media agent simulates conversation rather than collecting real social posts.

I create another project with dependencies and configuration, then define a custom get_ai_headlines tool. Its docstring describes the argument and the pipe-separated headlines it returns. The displayed code requests TechCrunch’s AI RSS feed, checks the response and handles errors; the tool retrieves headlines, not a verified body of independent research.

The social-media specialist receives a system prompt telling it to simulate online analysis, including sample posts and sentiment. I then describe research and other specialist agents through their roles, prompts and tools. Synthetic posts, percentages and engagement figures are examples generated by the model, not measurements of real conversations.

I describe an agents-as-tools approach, in which an agent can delegate to specialists, and note that Strands also offers other orchestration patterns. The project sets up interactive input: a user asks about a topic such as OpenAI, the news tool gets headlines, and the agents produce a summary. The demo illustrates a workflow, not proof that every generated claim is correct.

Custom tools are powerful because a developer can wrap programmatic functions for the agent to use. I do not explain each agent line by line, because the larger example builds on the earlier tool-and-prompt pattern. I say I will provide the code so viewers can inspect and customize the project.

The demonstration uses multiple Bedrock models, including Amazon Nova Pro, Nova Lite and Claude Sonnet 4. We added code to choose between models; I stress that the dynamic selection shown is our own application logic, not a feature automatically supplied by Strands. A developer could describe model strengths and design a routing procedure around the task.

In the main file, we import the tools, load configuration and run the specialists in the intended order before an expert produces the report. I choose a model and enter ‘Stargate and current US AI funding.’ The displayed interface distinguishes live AI news analysis from custom-topic analysis, so the input available depends on the selected path.

The demonstration shows the topic processor followed by social-media simulation, research, strategic analysis, sentiment analysis, recommendations and a final executive synthesis. It creates a two-page report. Although the screen displays figures such as a 20% funding increase and social sentiment percentages, the video does not verify those figures or establish that the simulated social analysis represents real observations.

I say the agents have shared memory and can use previous context. Memory and state sharing depend on the orchestration pattern and configuration; this demo does not establish automatic durable memory across every future conversation. Generated reports should be checked against identifiable sources before decisions rely on them.

I compare a graph to a flowchart, with dependencies connecting agent steps, and contrast it with a more exploratory swarm. This is a simplified explanation: current Strands documentation describes Swarm as autonomous handoffs with shared context, not inherently simultaneous parallel execution. Graph behavior, branching and parallelism depend on the SDK and graph configuration.

The choice of orchestration can be adapted to the task rather than fixed forever. I close by emphasizing the framework’s open-source code, support for different models and ability to integrate with other tools, and thank AWS for sponsoring the tutorial. Readers should consult the current SDK documentation and service pricing before reproducing the 2025 setup.