AWS Strands Agents Releases Strands Harness
AWS Strands Agents launches Strands Harness, an open-source agent harness delivering 28% lower token costs at comparable accuracy.
Stock photo for illustration only, not from the actual event
- AWS Strands Agents team launches Strands Harness open-source agent harness.
- Achieves 28% lower token costs compared to rivals at matching accuracy levels.
- Built-in defaults for prompt caching, context management, and local deployment.
- Installs easily via pip or npm with a dedicated CLI for automated prototyping.
The AWS Strands Agents team has officially released Strands Harness, a general-purpose open-source harness designed to serve as the supporting system around artificial intelligence models. Encompassing the loop, tools, context handling, memory, and recovery mechanisms, the package builds upon the existing Strands Harness SDK to bundle these essential building blocks into working defaults right out of the box.
Engineered as a general-purpose agent rather than a dedicated coding tool, Strands Harness allows developers to invoke create_harness() to instantly spin up agents equipped with bundled skills files. This enables coding agents to seamlessly generate deployment configurations for major cloud environments including AWS, GCP, Azure, Cloudflare, and Modal while running locally on standard workstations.
To validate performance, the team conducted distributed benchmarking on Amazon EC2 utilizing Harbor, the evaluation framework created by the authors of Terminal-Bench. Scores represent averages across six benchmarks: ALFWorld, ContextBench, GAIA, WebShop, τ²-bench, and Terminal-Bench 2.1, evaluated alongside competing systems like Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness.
Stock photo for illustration only, not from the actual event
Evaluation data highlights that while DeepSeek Harness proved to be the most token-efficient option overall—running roughly 14% cheaper than Strands harness—it trailed in accuracy across every benchmark. The highest-scoring data point on the chart belonged to Claude Opus 5 running on Strands harness at near 85%. In a direct head-to-head comparison on Terminal-Bench 2.1 across 89 trials, Strands harness cost 77% less than Claude Code while scoring 7.9 points higher in accuracy.
The ability of Strands Harness to drastically reduce token expenditure while preserving benchmark performance underscores a critical industry shift. As foundational models become increasingly capable, optimizing the surrounding harness layer—particularly through disciplined context handling and prompt caching—is proving to be the most effective strategy for controlling operational overhead.
Behind the scenes, the framework ships with robust defaults for context management and prompt caching, driven by a strict set of three operational rules. Furthermore, the accompanying Strands CLI allows developers to prototype complex agent workflows using plain English before exporting the resulting framework as a clean Python or TypeScript package.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment