AWS is releasing Strands harness, a ready-to-use agent harness licensed under Apache 2.0. According to AWS, the entire system runs with a single line of Python or TypeScript code, and costs 28 percent less than comparable harnesses, with no loss of accuracy.
Those working locally with Claude Code or Codex get an environment that does exactly what it’s supposed to do right away. If you build your own agent, you have to start over with prompts, tools, and context management. AWS aims to bridge that gap with Strands harness, which targets general-purpose agents, not just coding tasks.
The harness can be installed via ` pip install strands-harness ` or `npm install @strands-agents/harness`. It runs on models from Amazon Bedrock, Anthropic, OpenAI, and Google, as well as on local Ollama models or LiteLLM. Deployment is possible in any environment with a Linux container, from Cloudflare Containers and Google Cloud Run to Amazon ECS.
Fewer tokens, same scores
Across six benchmarks, Strands Harness used 28 percent fewer tokens when running the same Claude or GPT models. With Fable 5, the cost difference compared to Claude Code was as high as 77 percent, while achieving a higher score on Terminal Bench 2.1. Deepseek Harness was the most token-efficient, but according to AWS, it generally scored lowest on accuracy. Testing was conducted via distributed benchmarking on EC2 using Harbor.
The gains are primarily in the defaults. Tool results exceeding about 1,500 tokens are truncated; compaction starts as soon as the context window exceeds 85 percent; and on overflow, context recovery runs within the loop. Prompt caching is enabled by default. A paper by AWS researchers on these benchmarks is forthcoming.
Using the Strands CLI, developers can prototype an agent in natural language and export the code to Python or TypeScript using the `/export` command.
Tip: AWS can no longer recover some data from affected data centers