Strands Agents SDK

About

Strands takes a model-driven position: instead of orchestrating explicit workflows, the developer writes a few lines, the model chooses tools, and the SDK runs the loop with production controls around it — turn limits, token budgets, cancellation, stop-reason handling, structured output, streaming, session memory, and guardrails that catch mistakes before execution. Hooks intercept any step for logging, validation, or redirection, and steering handlers let an agent self-correct rather than fail; multi-agent patterns are first-class alongside a dedicated strands-mcp server package. First-class providers are Amazon Bedrock (default), Anthropic, OpenAI, and Gemini, with the SDK running entirely in-process and no hosted control plane. The Python and TypeScript SDKs share a monorepo with a docs site at strandsagents.com, and the project evolved from an internal AWS tool into a broadly used open-source framework.