AWS Strands Agents Team Releases Strands Harness: An Open-Source Agent Harness With 28% Lower Token Cost at Comparable Accuracy

The Strands Agents team at AWS released Strands harness: a packaged, general-purpose agent loop under Apache 2.0 for Python and TypeScript. It targets the familiar failure mode where a prototype works inside Claude Code or Codex, then falls apart when you rebuild the surrounding system yourself.

Harness choice is becoming an economic lever. Same model, wildly different spend. If the team’s benchmarks hold, defaults for caching and context management matter as much as which frontier model you pick.

We’re sympathetic to the problem statement. Agents aren’t the weights — they’re the loop, tools, memory, recovery, and how you truncate a 10,000-token tool dump before it lights your bill on fire. Shipping create_harness() with shell, file, and web tools, session resume, a helper agent, and Agent Skills support is useful infrastructure, not vapor.

The headline 28% lower cost across six benchmarks versus rivals — with near-equal accuracy — needs the footnote they included. DeepSeek Harness was cheaper still (~14% under Strands) but scored lower everywhere; including it pulled the savings figure down to 28%. On Terminal-Bench 2.1 with Claude Fable 5, Strands ran $56.29 at 69.7 accuracy versus Claude Code’s $248.05 at 61.8. That’s the chart that will get shared — and the one procurement teams will ask about.

What we like operationally: truncation above ~1,500-token tool results, compaction past 85% context, and in-loop recovery on overflow. That matches the broader HarnessTax finding they cite — harnesses barely move success rates while the same model can cost up to 5x more. Efficiency is a design choice, not a model lottery ticket.

Eyes open: MarkTechPost is amplifying a vendor release with vendor benchmarks on Harbor/EC2. Independent replication will matter. Still, open-sourcing a deployable harness that targets Bedrock, Anthropic, OpenAI, Google, Ollama, and LiteLLM is a pro-useful-AI move. The risk if everyone copies “smart defaults” poorly is silent context loss that looks like accuracy until a production edge case bites.

I’m watching the promised follow-up paper, whether the CLI-to-production export path stays honest, and how Strands compares once third parties re-run the six-benchmark suite.

Covered by MarkTechPost on Sep 21, 2026, summarizing the AWS Strands Agents team release. Benchmarks cited include ALFWorld, ContextBench, GAIA, WebShop, τ²-bench, and Terminal-Bench 2.1.