AWS Strands Agents SDK Enhances AI Agent Integration with Open Protocols

The Strands Agents SDK is an open-source, Apache 2.0-licensed SDK for building AI agents. It is model-driven, allowing developers to guide agent behavior through tools, prompts, and runtime configuration, and by adopting open standards, it enables agents to adapt to ecosystem changes.

MCP defines the contract for how agents access external tools, data, or context, answering the question of what resources are available. This allows models to interact with systems like grocery catalogs for actions such as checking prices or inventory without requiring hand-written integration code within the agent.

MCP support is built into Strands, shipping as a default dependency via an MCPClient. This client implements the ToolProvider interface, allowing it to be easily combined with other tools in the Agent constructor, enabling straightforward connection to external tools and services exposed by MCP servers.

A2A protocols facilitate complex workflows that require decision-making and judgment beyond simple tool calls. This pattern is useful for scenarios like finding product substitutions, where an agent might need to delegate tasks or engage in multi-turn conversations with other specialized agents.

The article uses 'Pantry,' a grocery delivery concierge, as a running example to demonstrate how protocols like MCP and A2A work together. Pantry uses MCP to access a grocery catalog for inventory checks and substitutions, showcasing how these protocols fit into a real system architecture.

A growing list of AI agent protocols can be overwhelming. Open source protocols like MCP, A2A, and UTCP are emerging to define standard boundaries and integration patterns for agentic systems, allowing components to interact without needing detailed knowledge of each other's implementations.