Strands Agents
★ 6.4k · GitHub →The model-driven agent SDK from AWS: an agent in a few lines, a full harness when you need control.
What is Strands Agents?
Strands Agents is the open-source SDK AWS uses to build its own agents — Amazon Q Developer and AWS Glue among them — released under Apache-2.0 and grown into a monorepo with both Python and TypeScript SDKs. Its founding bet is "model-driven": instead of hand-wiring orchestration graphs, you give a capable model a system prompt and a set of tools, and let the model plan the loop. An Agent is often three lines of code; complexity lives in the tools and the model, not the framework.
The "harness" framing is what distinguishes the current SDK: context management, execution limits, hooks that intercept any step of the loop, guardrails, and steering handlers that let an agent correct course instead of failing silently — all present before you write config. It is model-agnostic (Bedrock is the default, with first-class Anthropic, OpenAI, and Gemini support), speaks MCP natively, and covers multi-agent patterns, streaming, and structured output. Deployment pairs naturally with Amazon Bedrock AgentCore for managed runtime, memory, and identity.
We can speak to this one first-hand: this site's own backend runs Strands Agents on Bedrock — a 24-tool agent with skill packs, sub-agent spawning, and context compression via a summarizing conversation manager. The honest caveats: the community is younger and smaller than LangChain's, so you will read source code where others read Stack Overflow, and while it is genuinely multi-cloud, the smoothest path (Bedrock auth, AgentCore deployment) clearly runs through AWS.
Key capabilities
- ▸ Model-driven loop: define an agent with a prompt and tools in a few lines — the model plans, no graph wiring
- ▸ Full harness controls: hooks on every loop step, guardrails, steering handlers, context management, execution limits
- ▸ Python and TypeScript SDKs with native MCP support, streaming, structured output, and multi-agent patterns
- ▸ Model-agnostic: Bedrock by default, first-class Anthropic, OpenAI, and Gemini, plus custom providers
- ▸ Production path via Amazon Bedrock AgentCore (managed runtime, memory, identity) — battle-tested inside AWS products
When to use it
Pick Strands if you are on AWS or Bedrock — nothing else integrates as cleanly with Bedrock auth, AgentCore deployment, and the surrounding AWS stack — or if you buy the model-driven philosophy that modern models plan better than hand-built orchestration graphs. We use it for this site's production backend and the few-lines-to-working-agent claim holds. Skip it if you need a huge integration catalog (LangChain), explicitly checkpointed graph workflows (LangGraph), or if your team wants the larger community and tutorial base of the older frameworks.
Alternatives
Frequently asked questions
Is Strands Agents only for AWS?
No, though AWS is the smoothest path. The SDK is Apache-2.0 open source and model-agnostic, with first-class support for Anthropic, OpenAI, and Gemini APIs alongside the default Amazon Bedrock provider, plus custom providers. You can run a Strands agent anywhere Python or Node runs. What is AWS-specific is the managed deployment story: Bedrock AgentCore for runtime, memory, and identity.
What does "model-driven" mean in Strands Agents?
It means the model does the planning. Instead of defining an orchestration graph or state machine, you give the agent a system prompt and tools; the model decides which tools to call, in what order, and when the task is done. This keeps agent code tiny — often a few lines — and improves automatically as models improve, at the cost of less deterministic control flow than graph-based frameworks.
Does AWS actually use Strands Agents in production?
Yes. Strands originated inside AWS and powers agents in AWS products including Amazon Q Developer and AWS Glue — it was open-sourced after proving itself internally. It also pairs with Amazon Bedrock AgentCore as the recommended SDK for building agents on that platform. Independently, smaller teams (including this site's backend, a 24-tool agent on Bedrock) run it in production.
Star count updated 2026-07-02 · refreshed weekly from the GitHub API · browse all tools →