# Strands Agents Strands Agents is an open-source ([[Apache 2.0 License]]) production-ready AI agent SDK, built from production systems inside Amazon and maintained by the strands-agents org on GitHub. Its defining idea is a **model-driven approach**: rather than hard-coding orchestration, you give the model tools and let an [[Strands Agent Loop|agent loop]] drive reasoning, tool selection, and execution. It ships in [[Python]] (`strands-agents`) and [[TypeScript]] (`@strands-agents/sdk`) with near-identical APIs. ## TypeScript status The [[TypeScript]] SDK (`@strands-agents/sdk`) reached **1.0.0 on 2026-04-30** and has moved quickly since — **1.13.0** as of 2026-08-20 (1.13 shipped 2026-08-12). Its first public release was 0.1.0 in December 2025, so the TS line is young: about eight months old, four of them post-1.0. Two things follow from that. The TS API is stable enough to build on but the community around it is small compared to the Python line, so expect fewer worked examples and slower answers. And Strands remains **Python-first on features** — new capabilities tend to land there first and reach TS on a lag, so verify a given feature exists in the TS docs rather than assuming parity. The AWS coupling is softer than the Amazon origin suggests: Bedrock is the default model provider, not a requirement, and provider plurality is achieved partly through [[Vercel AI SDK]] compatibility. File-based session persistence is a notable design choice if you want sessions that live as inspectable files rather than rows in a database. ## Pillars - **Flexibility**: any model, any cloud (14+ official model providers: Bedrock, [[Anthropic]], [[OpenAI]], Google, [[Ollama]], Mistral, Cohere, …) - **Control**: [[Strands Hooks|hooks]], guardrails, and steering handlers for policy enforcement and self-correction - **Production-readiness**: observability, [[Strands Context Management|context management]], session persistence, and a [[Strands Agents Shell|sandboxed shell]] ## Core concepts - **[[Strands Agent Loop]]**: the recursive reason → select tool → execute → feed back cycle - **Tools**: built-in (community library), custom (`@tool` decorator), or [[Model Context Protocol (MCP)|MCP]] tools; executors run them concurrently or sequentially - **[[Strands State]]**: multi-layer state (conversation history, agent/app state, invocation state) - **[[Strands Context Management]]**: automatic or agentic strategies to manage the token budget - **Session management & memory**: persist across sessions (file or S3 backends; recall/inject/extract memory) - **Interrupts**: human-in-the-loop pauses; **multi-agent** orchestration via Graphs, Swarms, and A2A ## References - https://strandsagents.com/ - https://strandsagents.com/docs/user-guide/quickstart/typescript/ - https://strandsagents.com/blog/ - https://github.com/strands-agents ## Related - [[Strands Agent Loop]] - [[Strands Hooks]] - [[Strands State]] - [[Strands Context Management]] - [[Strands Agents Shell]] - [[AI Agents]] - [[Agentic Engineering]] - [[AI Agent Harness]] - [[Model Context Protocol (MCP)]] - [[AI SDKs]] - [[TypeScript]] - [[Python]] - [[Vercel AI SDK]] - [[Apache 2.0 License]]