AWS Open-Sources Strands Harness Agent Framework

AWS introduced Strands Harness on September 21, 2026: an open-source agent harness that ships the loop around a language model rather than leaving developers to assemble it — tool calls, memory, context handling and error recovery. It is licensed under Apache 2.0 and comes in a Python and a TypeScript build.

Harness versus SDK

An SDK hands you parts; a harness hands you the wiring. Strands Harness arrives with shell, file access — read, write, edit — and web search already configured. On top of that sit long-term memory addressed by session ID, a helper agent that works open-ended subtasks off a checklist, and direct support for Model Context Protocol servers. Both reports state that any default can be overridden, down to the SDK layer itself.

Models and deployment targets

Amazon Bedrock, Anthropic, OpenAI, Google, Ollama and LiteLLM can all serve as the model backend, so locally hosted Ollama models are as much in scope as frontier models in the cloud. The harness runs on a developer machine and deploys to AWS, Google Cloud, Microsoft Azure, Cloudflare and Modal. MarkTechPost additionally describes a Strands CLI for sketching an agent in plain language and exporting the result as a Python or TypeScript package.

What the benchmark numbers claim

The cost figures are doing most of the work in this announcement. On Terminal-Bench 2.1, AWS puts Strands Harness at 77 percent below Claude Code — $56.29 against $248.05 — while scoring higher, 69.7 percent versus 61.8 percent, a gap of 7.9 points. SiliconANGLE attributes that comparison to a run on Anthropic Fable 5, while MarkTechPost cites a best result of close to 85 percent with Claude Opus 5 on the harness.

Where the two reports diverge

On the wider comparison the accounts do not line up. MarkTechPost reports 28 percent lower cost than rival harnesses across six benchmarks — ALFWorld, ContextBench, GAIA, WebShop, τ²-bench and Terminal-Bench 2.1 — at near-equal accuracy, whereas SiliconANGLE reports 26 percent greater efficiency on the same underlying model. Which of the two matches the original AWS figure cannot be settled from these two texts. Every number here is vendor-measured and has not been independently reproduced so far.

Check the package name before installing

The install instructions also disagree: MarkTechPost lists strands-harness for pip and @strands-agents/harness for npm, while SiliconANGLE gives strands-agents-harness for both. Confirm the name against the project page rather than either write-up. The team credits context management as the actual efficiency lever — bulky tool results are written out to separate files instead of filling the context window, combined with prompt caching.