SEATTLE, September 22 — Amazon Web Services has released an open-source AI agent it says can run just about anywhere, a move aimed at the stubborn gap between a prototype that works on a laptop and an application that has to scale. The company announced Strands Harness on September 21, 2026, in a blog post published that day. AWS describes it as a complete, ready-to-use agent that runs locally or in any cloud — not only AWS but also Google Cloud, Microsoft Azure, Modal, Cloudflare and others.
Here’s the problem it targets. Plenty of developers have already built AI agent prototypes on their own machines using tools such as Anthropic’s Claude Code and OpenAI’s Codex, because those environments “just worked” locally.
Moving that work into a scalable cloud setup is where things get difficult. Strands Harness is AWS’s attempt to smooth that transition.
What’s inside the box
The agent is built on top of the Strands Harness software development kit, an open-source framework AWS developed for managing and running multi-agent patterns. AWS said the framework can run on top of the latest frontier models from Anthropic, OpenAI, Amazon Bedrock and Google. Developers who would rather keep their agents hosted locally can point Strands Harness at an Ollama model running on their own machine.
AWS said the agent ships with essential components out of the box, including read, write and edit, shell and web search capabilities. Rather than asking developers to design bespoke tools for every agentic task, Strands Harness leans on tools the underlying model already knows how to use.
Think of it like a fully stocked kitchen rather than an empty countertop. The point is that developers spend their time on the application, not on assembling basic utensils first. The agent also manages its own context window, offloading tool results to separate files and caching reused parts of requests.
AWS said that cuts processing times and makes the agent more efficient in terms of token consumption. Strands Harness keeps a long-term memory across runs, so it can resume previous conversations through session IDs.
A built-in helper agent can take on open-ended subtasks using an automated checklist. Developers can upload various “Agent Skills” and integrate external tools such as Model Context Protocol servers.
The numbers AWS is citing
AWS said Strands Harness performed strongly across a range of industry benchmarks. The company said the agent was 26% more efficient than agents built on other frameworks when using the same underlying model. It also reported greater token efficiency than Claude Code and Codex, and said Strands Harness beat those agents on several other benchmarks.
In one test using Anthropic’s Fable 5 model, AWS said Strands Harness cost 77% less than Claude Code on the same tasks while achieving a higher overall score on the Terminal Bench 2.1 benchmark. AWS said the primary use case is rapid prototyping of new AI applications.
One of its own teams used Strands Harness to build the Strands CLI, a command-line interface that lets developers — and people without coding skills — prototype AI agents using natural language commands. A user selects the underlying model, adds the prompts and tools they think the agent needs, then runs an “/export” command to download the underlying harness code as a Python or TypeScript file. Strands Harness is available to download from GitHub.
It can also be installed through standard package managers with pip install strands-agents-harness or npm install strands-agents-harness. The practical upshot for anyone building agents at home or at work is a shorter path from a weekend prototype to something that runs in production.
What to watch next is whether developers outside AWS adopt it at the same rate the company’s own teams have, and whether the benchmark advantages hold up once other frameworks respond.
