AWS bolts together open source agent harness, says it sips fewer tokens than rivals

Strands claims near-parity with Claude Code and Codex on benchmarks, though it only raced coding agents and marked its own homework

AWS has entered the open source agentic AI game, claiming its new Strands harness matches rivals on benchmarks while using around a quarter fewer tokens.

The Strands harness, as its name suggests, is built on AWS’ Strands Harness SDK, but is packaged up and ready to roll out of the box, either locally or deployed to work with whatever AI provider a customer prefers.

"The moment you build your own agent, you're on your own," AWS noted in its announcement. "It's tricky wiring up the right primitives just well enough to match that ‘it just worked’ feeling."

REG AD

This new open-source option means that 'all you need is one line of Python or TypeScript to get it working with your choice of model," AWS adds. 

REG AD

On top of being relatively plug-and-play in design, AWS claims the Strands harness achieved "nearly equal benchmark scores" versus Claude Code, Codex, and “other popular harnesses” when tested using the Harbor framework, distributed across multiple nodes of AWS’ own EC2 virtual servers for benchmarking tests.

Strands consumed 28 percent fewer tokens across six benchmark tests when compared to “Claude or GPT models,” claims AWS, and in some cases had better accuracy than other harnesses too. DeepSeek's harness, AWS admits, was more token efficient but - we're told - was less accurate on the same tests.

AWS credits this performance to the Strands harness’ default prompt caching and context management settings. Per the announcement post, it defaults to truncating tool results over 1,500 tokens, automatically compacting its context window when it surpasses 85 percent, and automatically trying to recover context in overflow cases.

It’s worth pointing out, however, that AWS describes the Strands harness as being built to be "a general-purpose agent rather than a coding agent." Considering that every agent harness it was compared to in AWS’ benchmarks (Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness) is billed as a coding agent, it’s worth questioning how other general-purpose AI agents would perform when compared to the new Strands harness.

Amazon plans to publish a paper from its researchers with more benchmark details - we’ll keep an eye out to see more of the specifics on how it was evaluated. 

In the meantime, anyone who wants to give the pre-packaged, open-source, general-purpose Strands harness a shot can do so by using the instructions included in the announcement post. It’s available as a Python or TypeScript library, and also has a CLI. And it ships with a skills file, so plugging in your own preferred tools should be relatively easy; shell, file, and web tools are included at launch. 

“We believe quickly prototyping with a batteries included harness will create more useful agents,” AWS says of the Strands harness design. It’s live now, so if you’re looking for another open-source option it’s ready when you are, even if a more independent gauge of its performance is currently lacking. ®