Strands Harness

Claude Code and Codex setups often ran locally, but when building agents, it became challenging to replicate the "it just worked" experience. Strands harness, a new fully assembled, state-of-the-art agent harness, is now available for easy local deployment or provider deployment. It's a general-purpose agent that can be run with a single line of Python or TypeScript code, supporting Claude or GPT models.

Strands harness outperforms other agent harnesses in cost efficiency while maintaining equal or better accuracy, with 28% lower costs across six benchmarks. It also offers better token-efficiency and similar benchmark scores compared to Claude Code, Codex, and other popular harnesses. Deepseek Harness, however, showed the highest token-efficiency but lower accuracy.

Strands harness incorporates defaults for prompt caching and context management, which contribute to its token-efficiency and accuracy. It's compatible with Amazon Bedrock, Anthropic, OpenAI, and Google models, and can be deployed on any provider with a Linux container. A Strands CLI allows users to prototype agents in plain English and export the code in TypeScript or Python.

The harness is fully customizable, allowing users to swap models, add tools, or replace components. Installation is simple with pip or npm, and the team is available on Discord for further assistance.