AWS's Open-Source Strands Harness Undercuts Claude Code 77%

AWS launched Strands Harness, an open-source AI agent claiming 26% better efficiency and 77% lower cost than Claude Code on the same Anthropic model.

Amazon Web Services today launched Strands Harness, an open-source agent that AWS says helps developers bridge the hardest step in agentic development: moving from a prototype that 'just worked' locally with tools like Claude Code or Codex to a deployment that scales in the cloud. Built on the open-source Strands Harness SDK for multi-agent patterns, it runs on frontier models from Anthropic, OpenAI, Amazon Bedrock and Google, or on a local Ollama model. It ships with read, write, edit, shell and web-search capabilities out of the box, relying on tools the underlying model already knows rather than bespoke per-task tools. It is downloadable from GitHub or installable via pip or npm.

The surprise is the benchmark claim. According to SiliconANGLE, AWS reports Strands Harness was 26% more efficient than agents built on other frameworks using the same underlying model, with better token efficiency than both Claude Code and Codex. In one test using Anthropic's Fable 5 model, SiliconANGLE reports that Strands Harness cost 77% less than Claude Code on the same tasks while scoring higher overall on the Terminal Bench 2.1 benchmark. An open-source harness from the cloud provider beating the incumbent coding agents on cost and score, on the incumbent's own model, is the kind of result agent builders will be pressure-testing this week.

SiliconANGLE attributes the efficiency to architecture rather than model choice. Strands Harness manages its own context window intelligently, offloading tool results to separate files and caching reused parts of requests to cut processing time and token consumption. It maintains long-term memory across runs, resuming conversations via session IDs, and can delegate open-ended subtasks to a built-in helper agent using an automated checklist. Developers can upload 'Agent Skills' and integrate external tools such as Model Context Protocol servers. AWS's own team used it to build Strands CLI, which lets non-coders prototype agents in natural language and export the harness code as Python or TypeScript.

The launch lands alongside AWS's broader open-model push. Per the South China Morning Post, Moonshot AI's Kimi K3 is now available on Amazon Bedrock, joining dozens of open-weight models from DeepSeek, MiniMax and Alibaba on the platform, and SCMP reports Moonshot was previously in talks with Microsoft and Google over revenue-sharing deals seeking as much as 30% of revenue. AWS told SCMP that customer data sent to Kimi K3 stays within its cloud security boundary with zero data retention for inference. A free, portable agent harness that runs on any cloud — plus Bedrock's expanding open-weights catalogue — gives AWS a developer-funnel story that competitors' closed agent tooling does not match.