An agent SDK that treats models as interchangeable parts

What it does Strands Agents gives you a minimal agent loop and tool system in Python and TypeScript. You instantiate an Agent, load tools via decorators or MCP servers, and swap the underlying model by changing a single provider argument. The SDK abstracts conversation loops, tool execution, and streaming across roughly a dozen hosted and local inference options.

The interesting bit The monorepo includes a WebAssembly bridge that lets TypeScript agents call Python tools, which solves the classic two-language problem without maintaining separate tool libraries. It also experiments with bidirectional streaming for real-time voice and audio against Nova Sonic, Gemini Live, and OpenAI Realtime.

Key highlights

• Native support for MCP servers, giving agents access to external tool ecosystems without custom glue
• Built-in model providers spanning Bedrock, Anthropic, Gemini, LiteLLM, Ollama, OpenAI, SageMaker, Writer, and custom implementations
• Optional strands-agents-tools package with pre-built utilities for quick prototyping
• Hot-reloading of local tool directories during development
• Experimental bidirectional streaming with interruptible, persistent audio conversations

Caveats

• The default provider path assumes AWS Bedrock credentials and Claude 4 Sonnet access in us-west-2, so the quickest quickstart is not cloud-agnostic
• Bidirectional streaming is explicitly experimental and the APIs may shift in future releases

Verdict A solid fit if you want one harness that moves from local Ollama experiments to production Bedrock without framework bloat. Less compelling if you need deeply opinionated multi-agent orchestration—the README advertises the capability but keeps the surface area thin.