AWS opens its agent harness to everyone under Apache 2.0

AWS hands developers the harness it uses to cut agent bills

AWS has published Strands harness, the agent scaffolding its Strands Agents team built for its own products, under an Apache 2.0 licence. The release lands on 21 September 2026 with a blunt commercial claim attached: on the same Claude or GPT models, across six benchmarks, the harness costs 28% less to run than other popular harnesses while scoring nearly the same.

The distinction that matters is what kind of agent this is. Strands harness is built as a general-purpose agent rather than a coding agent, so it ships with shell, file and web tools, keeps long-term memory between runs, and resumes an earlier conversation when handed a session ID. You start it with one line of Python or TypeScript, pointing the model string at Amazon Bedrock, Anthropic, OpenAI, Google, a local Ollama model or LiteLLM.

The claim, and who is making it

AWS says the 28% saving holds across six benchmarks at equal or better accuracy, that the harness cost 77% less than Claude Code when both ran Fable 5, and that it scored higher on Terminal Bench 2.1. It also reports that DeepSeek Harness was the most token-efficient harness tested but typically posted the lowest accuracy scores. The testing was distributed on EC2 with Harbor, and the company says a follow-up paper from its researchers is still to come - which means every figure here is the vendor marking its own homework until that paper lands.

Where the cost saving comes from

The mechanics are unglamorous and entirely about context. Tool results longer than roughly 1,500 tokens get truncated, bulky output is offloaded to files instead of being held in the window, reused parts of each request are cached, summarisation triggers once the context window passes 85%, and context recovery runs inside the loop if an overflow happens anyway. None of that is novel on its own; the pitch is that these defaults arrive switched on.

What you get, and where it runs

Beyond the tools, the harness ships with a built-in helper agent for open-ended subtasks, checklist tracking for multi-step work, skill loading, and deployment on any provider with a Linux container - Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS and Amazon Bedrock AgentCore are all named. The Strands CLI is built on the harness, and its export command hands back the agent as TypeScript or Python to keep iterating in a preferred coding tool.

The code is genuinely open. The harness-sdk repository is licensed Apache-2.0 and, as of 21 September 2026, showed 7,412 stars and 1,171 forks, with its description promising an SDK for production agents in Python and TypeScript against any model and any cloud.

What is not settled

The benchmark comparison is self-reported and the paper is unpublished, so the accuracy parity claim cannot be checked outside AWS yet. Nobody has published an independent run against the same six-benchmark suite, and the cost figures depend on prompt caching being available on the model you choose - which is not true everywhere. The harness also truncates tool results by default, and a truncated result is precisely the sort of thing that produces confident wrong answers in agent workflows, so the accuracy numbers are the ones to watch when the paper arrives.

Our opinion

The harness layer is where the money quietly moves, and AWS giving its own away is an admission that context management - not model quality - is now the differentiating engineering. The defaults are the interesting artefact here, because they encode a bet: that most agents are drowning in tool output they never needed, and that throwing away anything past 1,500 tokens buys more than it costs. An open-source release also means those defaults are now a baseline other vendors have to argue with rather than a feature they can sell.

The caution is that a cheaper agent is only cheaper if it still finishes the job. Cost per token is easy to publish and pleasant to quote; success rate per task is the number a team actually pays for, and it is the number the unpublished paper has to produce before anyone should rebuild their stack around a 28% claim. Until then this is a well-engineered, freely licensed starting point with an attractive invoice and an unproven finish.