Three Clouds, One Brief: What Actually Differs Between ADK, Strands and Agent Framework

All three major cloud providers – Google, AWS, and Microsoft – now offer an agent framework for managing language models. These frameworks, known as ADK, Strands, and Agent Framework respectively, adhere to the A2A (Agent-to-Agent) protocol which ensures interoperability between different agent implementations. Although A2A provides a common language for agent communication, the wire (data transmission) is only one part of the overall process.

To test the differences between these three agent frameworks, a researcher built the same agent three times – one on Google ADK, one on AWS Strands, and one on Microsoft Agent Framework – each hosted on their respective runtimes. The study used a single coordinator to send the same brief to all three agents and then scored the results. Google provides a ready-made search tool, whereas AWS Strands and Microsoft Agent Framework do not.

When the search tool was replaced with a generic one, the only variable left was the agent framework itself. The researcher examined the construction of each framework, noting that Google ADK uses the LlmAgent class, AWS Strands uses the Agent class with a BedrockModel object, and Azure Agent Framework uses the Agent class with a FoundryChatClient and DefaultAzureCredential. The system prompt, model specification, tools, and other parameters varied among the frameworks.

The main takeaway from this comparison is that while A2A ensures interoperability, the actual performance and efficiency of the agent heavily depend on the underlying framework. The choice of framework determines the shape of the agent and the way it interacts with the models and tools. Thus, when evaluating the performance of language models across different platforms, it is crucial to consider not just the A2A protocol, but also the specific agent framework used.