Strands Agents releases open-source AI agent harness with token-efficiency claim
The company says Strands Harness used 28% fewer tokens across six benchmarks while delivering similar or better accuracy. The comparisons have not been independently assessed in the supplied material.
Strands Agents has released Strands Harness, an open-source, customisable framework for building general-purpose AI agents that can run locally or in cloud environments. It supports Python and TypeScript and is licensed under Apache 2.0.
The company says the harness used 28% fewer tokens than competing systems across six benchmarks while achieving similar or better accuracy. Strands says its tests compared performance using the same Claude or GPT models and included Claude Code and Codex.
The benchmark methods and detailed results have not been independently assessed in the supplied material. Strands says it plans to publish a follow-up research paper; the company’s reported results should be read in that context.
The framework includes defaults for prompt caching and context management. These truncate tool results longer than about 1,500 tokens, trigger summarisation when context use exceeds 85%, and attempt recovery when the context window overflows.
Strands Harness works with models from Amazon Bedrock, Anthropic, OpenAI and Google, as well as local Ollama models. Strands says it can be deployed in Linux containers across cloud providers, and that users can customise its defaults, models and tools.
