Strands harness claims 28% lower token cost and credits its context defaults

Strands harness is a general-purpose agent, not a coding agent, released under Apache 2.0 for Python and TypeScript. It runs on Bedrock, Anthropic, OpenAI, Google, Ollama or LiteLLM and deploys to any Linux container host. The team reports that across six benchmarks, run on EC2 with Harbor, it cost 28% less than Claude Code, Codex and other harnesses on the same Claude or GPT models, with "nearly equal benchmark scores." With Fable 5 on Terminal Bench 2.1, it says, the harness cost 77% less than Claude Code and scored higher. Deepseek Harness used the fewest tokens but, by the post's account, usually scored lowest.

The post attributes most of the gain to defaults: tool results over about 1,500 tokens are truncated, compaction triggers above 85% of the context window, and context recovery runs inside the loop when the window overflows. It promises "a follow-up paper from our researchers regarding these benchmarks."

On Hacker News, where the post had 80 points and 51 comments about two hours after submission, Albert Zhao, posting as one of the authors, said the Strands team is part of AWS and gave the Terminal Bench 2.1 scores behind the chart: 69.7 for Strands against 61.8 for Claude Code, both on Fable 5 at high effort. Commenter seizethecheese called Terminal Bench 2.1 "saturated" and pointed to Artificial Analysis scores near 90% on it. Several others asked why the comparison included oh-my-pi but not plain pi; Zhao said the team can look into a pi run.

Why it matters: Truncating large tool results, compacting before the window fills and recovering inside the loop are settings any harness owner can copy without adopting this harness. Until the paper is out, treat the 28% figure as a vendor claim on vendor benchmarks and check the defaults against your own token bills.