AWS Open-Sources Strands Harness, a Cheaper Ready-Made AI Agent

What AWS announced

On September 21, 2026, AWS released Strands harness, an open-source, fully assembled AI agent that developers can run locally or deploy to any provider. Where a bare model needs someone to wire up the surrounding 'harness' of tools, context handling and memory, Strands ships those defaults in the box, and AWS pitches it as a general-purpose agent rather than a coding-specific assistant. It is available under a permissive Apache 2.0 license and takes a single line of Python or TypeScript to start.

The cost claim, and the caveats

AWS's headline is efficiency: it says Strands harness costs 28% less than Claude Code and Codex when using the same Claude or GPT models across six benchmarks, while scoring roughly equal or better on accuracy. The framing in the wider coverage, including The New Stack, is that a better harness can matter more than a better model. There are honest caveats, though. AWS itself notes that DeepSeek Harness proved the most token-efficient overall, running cheaper than Strands, but typically posted the lowest accuracy scores. In one configuration, using a model AWS calls Fable 5, Strands said it cost 77% less than Claude Code while scoring higher on Terminal Bench 2.1.

Why it is cheaper: context management

AWS attributes most of the savings to how Strands handles the context window rather than any single trick. Its defaults include prompt caching plus context management that truncates tool results over roughly 1,500 tokens, triggers summarization (compaction) once the context window passes about 85% full, and runs context recovery inside the loop if there is an overflow. Offloading bulky tool output to files and caching the reused parts of each request is what drives the token efficiency, AWS says, with a benchmarks paper promised as a follow-up.

Model-agnostic by design

Crucially for anyone wary of lock-in, Strands is not tied to AWS. It runs on current models across Amazon Bedrock, Anthropic, OpenAI, Google, Ollama and LiteLLM, and you select one by name or point it at a local Ollama model. Out of the box the agent ships with shell, file (read/write/edit) and web tools, manages its own context window, keeps long-term memory across runs (resuming via a session ID), delegates open-ended subtasks to a built-in helper agent, and can load skills. It can be deployed to any provider that runs a Linux container, including Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run and Amazon ECS.

The bigger picture

Strands lands in a busy moment for agent 'harnesses,' the connective tissue around models that a wave of tools, from Claude Code to Codex to a growing set of open-source rivals, are all competing to own. By giving away a batteries-included, model-agnostic harness under Apache 2.0, AWS is betting that lowering the barrier to a good default setup, and the token bill that comes with it, will pull more developers into building agents its way. Whether AWS's efficiency numbers hold up under independent testing is the open question its promised benchmark paper will need to answer.

Related on Skillo

See also: Meta's Llama 3 open models, Llama 3.2's vision and edge models.

Sources

• Introducing Strands harness (Strands Agents official blog, 21 Sep 2026)
• AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex (The New Stack)

Published date reflects the original event date (2026-09-21). This article is original Skillo editorial written from the sources above; facts were verified in September 2026.