Strands Agents

Strands Agents is an SDK for building and running agents in Python and TypeScript. It runs inside your own process with no hosted control plane, and is positioned for the case where you would otherwise write the agent loop yourself: it covers the parts such a loop grows into rather than prescribing an application architecture.

In one SDK it provides lifecycle controls — turn limits, token budgets, cancellation and stop reasons — tools and structured output, Model Context Protocol (MCP) support, multi-agent patterns, memory and session management, model portability, streaming, guardrails, tracing and evaluations. Hooks let you intercept any step of the loop to log it, validate it or redirect it, and steering handlers let an agent correct itself rather than failing silently.

Model support is first-class for Amazon Bedrock, Anthropic, OpenAI and Gemini, with further providers including Ollama and the option of writing a custom one; both SDKs default to the Bedrock provider. Swapping backends is intended to leave application code unchanged.

The project is a monorepo holding the Python SDK (Python 3.10 and later), the TypeScript SDK (Node.js 20 and later), the documentation site, and cross-SDK governance material covering tenets, decisions and compatibility guidelines. Separate repositories carry samples, a tools package and an MCP server.

Features

• Agent loop: a model-driven loop with turn limits, token budgets, cancellation and explicit stop reasons
• Tools and structured output: Python and Zod-typed TypeScript tool definitions, plus typed results from a run
• MCP: Model Context Protocol tools are usable directly from an agent
• Multi-agent patterns: built-in patterns for coordinating several agents
• Memory and sessions: session management and memory as first-class SDK concerns
• Model portability: Amazon Bedrock, Anthropic, OpenAI, Gemini, Ollama, further providers and custom ones behind the same code
• Hooks: intercept any loop step to log, validate or redirect it
• Guardrails and steering: checks that stop a bad action before it runs, and handlers that let an agent correct itself
• Observability: the loop traces every decision by default, with tracing and evals in the same SDK
• Two languages: matching Python and TypeScript SDKs maintained in one repository