Skip to content

Strands Agents

Strands Agents, the open-source, model-driven agent SDK from AWS, in Python and TypeScript: what it is, how the two SDKs compare, when to choose it, and how it relates to Amazon Bedrock and AgentCore.

GroupCloud-Native
Current versionPython v1.64.0 / TypeScript v1.24.0
Bonds withPyTs

Strands Agents is an open-source (Apache 2.0) SDK from AWS for building agents. It is model-driven: you give an agent a model, a system prompt and tools, and the model plans the steps and decides when it is done. The SDK supplies the loop around that: tool execution, conversation management, sessions, hooks and policies, multi-agent orchestration, MCP and A2A, and OpenTelemetry tracing. It uses Amazon Bedrock by default and works with Anthropic, OpenAI, Google and other providers.

Current versions: Python strands-agents 1.64.0 with strands-agents-tools 0.9.0 · TypeScript @strands-agents/sdk 1.24.0 (checked 2026-11-19).

Python install

Terminal window
pip install "strands-agents==1.64.0"

TypeScript install

Terminal window
npm install @strands-agents/sdk@1.24.0 zod
from strands import Agent, tool
@tool
def get_weather(city: str) -> str:
"""Get today's forecast for a city.
Args:
city: City name.
"""
return f"Light rain in {city}, 14°C"
agent = Agent(model="us.anthropic.claude-sonnet-5-5", tools=[get_weather])
agent("Do I need an umbrella in Leeds today?")

There is no graph to draw and no prompt template to fill: the model sees the tool’s schema, decides to call it, reads the result and answers. Everything else in the SDK is about controlling that loop: bounding it, observing it, persisting it, pausing it for a human, and combining several agents.

PythonTypeScript
Packagestrands-agents (+ strands-agents-tools, strands-agents-evals)@strands-agents/sdk
RuntimePython 3.10+Node.js 22+, ESM only
Tools@tool functions (schema from type hints and docstrings)tool() with Zod schemas, fully typed
ProvidersBedrock, Anthropic, OpenAI, Gemini, LiteLLM, Ollama, Mistral, llama.cpp, Llama API, SageMaker, WriterBedrock, Anthropic, OpenAI, Google, Vercel AI SDK
Multi-agentAgents as tools, Swarm, GraphBuilder, model-driven graph tool, A2AAgents as tools, Swarm, Graph, A2A
ExtrasEvals package, real-time voice (BidiAgent), the larger tool libraryVended tools only

Both SDKs share the concepts and most names, so moving between them is mostly syntax. Choose by the language of the service the agent lives in; choose Python if you need evals, the wider tool set or a provider only Python supports.

Good fit:

  • Tool-using agents where the model should plan the steps, rather than a fixed workflow you would draw as a graph.
  • Agents on AWS: Bedrock models, IAM credentials, AgentCore Runtime hosting, AgentCore Memory and Gateway all fit without glue.
  • Teams that want one SDK across providers, with the option of Python or TypeScript.
  • Multi-agent systems that start simple (a supervisor with specialists as tools) and grow into swarms or graphs.

Consider something else when:

  • You need explicit, checkpointed state machines with time travel and branching as the core design: LangGraph is built around that.
  • You are committed to OpenAI’s platform features (hosted tools, Realtime, tracing dashboard): the OpenAI Agents SDK fits more naturally.
  • You want a fully managed agent with no code at all: AgentCore’s managed harness or Bedrock Agents (see the Amazon Bedrock guide).

These are three different things that work well together:

PieceWhat it isWhere it is covered
Strands AgentsAn SDK: the agent loop runs in your processThis guide
Amazon BedrockThe model service; Strands’ default provider (BedrockModel), plus Guardrails and Knowledge BasesModel providers
Amazon Bedrock AgentCoreManaged infrastructure for agents written with any framework: Runtime (hosting), Memory, Gateway (tools as MCP), Identity, Code Interpreter, Browser, ObservabilityThe Amazon Bedrock guide

A typical AWS deployment is a Strands agent inside a BedrockAgentCoreApp on AgentCore Runtime, calling Bedrock models, keeping conversations in AgentCore Memory and reaching internal APIs through AgentCore Gateway. Strands does not require AgentCore, or even AWS: the same agent runs in Lambda, a container, or on your laptop against the Anthropic or OpenAI API. Hosting is covered on the production pages (Python, TypeScript).


DateVersionChanges
2026-10-24Py 1.64.0 / TS 1.24.0New standalone guide. Python and TypeScript sections with comprehensive, tools and MCP, hooks and interrupts, multi-agent and A2A, observability and evals (Python), recipes and production pages, plus shared diagrams. Replaces the Strands page in the Amazon Bedrock guide; all examples run against the installed packages with an offline Bedrock stub.