Strands Harness SDK

Strands Agents is an open-source SDK from Amazon Web Services for building and running AI agents in Python and TypeScript. The harness-sdk monorepo contains the assembled Strands harness, the Python and TypeScript SDKs, a CLI, an MCP server for coding agents, and the documentation site. It runs in your own process with no hosted control plane, according to the README.

What It Is

The Strands Harness SDK is an agent framework for developers who would otherwise write their own agent loop. It runs the loop and provides components to attach to it: tools, model providers, memory, sessions, hooks, and context management. Alongside it, the Strands harness offers a fully assembled agent through a single create_harness() (Python) or createHarness() (TypeScript) call, with defaults for the system prompt, tools, memory, sessions and context management. Those defaults can be overridden or replaced.

How the Workflow Fits Together

Developers can start with the assembled harness, using the library or the strands CLI, then drop down to the SDK to control the loop directly. Functions become tools through a @tool decorator (Python) or a schema-based tool() helper (TypeScript), and MCP servers can be connected. The SDK documents lifecycle controls such as turn limits, token budgets and cancellation, structured output, streaming, and multi-agent patterns including Agents as Tools, Swarm, Graph, Workflow and Agent2Agent. Hooks, interventions (human in the loop, Cedar authorization, steering) and guardrails let developers intercept or redirect steps.

Models, Deployment and the Wider Toolkit

The SDK is model agnostic, with documented providers including Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, LiteLLM, llama.cpp, Mistral and SageMaker, plus custom providers. Deployment guides cover Amazon Bedrock AgentCore, Lambda, Fargate, App Runner, EKS, EC2, Docker, Kubernetes and Terraform. Observability covers metrics, traces and logs. The same site documents companion projects: Box (an agent sandbox, in preview), Shell (a sandboxed virtual shell), Evals (agent evaluation, Python only) and Labs (experimental projects).