AWS Strands Agents open-sources Strands harness general-purpose agent

The Strands Agents project released Strands harness, a fully assembled open-source AI agent you can run locally or deploy on any cloud, under an Apache 2.0 license, with one-line Python or TypeScript setup.

SOFTWARE desk — AWS’s open agent project is shipping a batteries-included harness, not just an SDK, aimed at the Claude Code / Codex local-to-cloud gap.

Why the post says the harness exists. Builders, it says, often wished a Claude Code or Codex setup could run in the cloud, because locally the agent already “just worked.” Claude Code is Anthropic’s coding agent. Codex is OpenAI’s. The moment you build your own agent, the post says, you are on your own, wiring the pieces until it feels that solid. Strands harness is the project’s answer: one line of Python, or one line of TypeScript, pointed at the model you choose. Python and TypeScript are two programming languages. File that framing as the project’s.

How you install it, still the post. Three commands are named. For Python: pip install strands-harness. For TypeScript: npm install @strands-agents/harness. npm is the tool that installs JavaScript packages. For an interactive command line: npm install -g @strands-agents/cli. A command line is the text window where you type. The same page says that command-line tool lets you sketch an agent in plain English, then run a command it calls /export to get a copy of the code in TypeScript or Python, so you can keep editing it with whatever coding agent you already use. File those three install paths and the export step as the post’s. This desk did not run them.

The 28 percent line is a company benchmark, not a score this desk ran. The post says Strands harness costs 28 percent less when using the same Claude or GPT models across six benchmarks. Claude is Anthropic’s model family. GPT is OpenAI’s. A benchmark is a fixed test used to compare tools. The post says the harness showed better token efficiency and nearly equal scores against Claude Code, Codex, and other popular harnesses. A token is a small chunk of text the model reads or writes, and you pay by the chunk. Token efficiency means fewer chunks for the same job. The post says the point is to keep accuracy while the bill drops. It also says a harness it calls Deepseek Harness was the most token-efficient overall, and that it typically reported the lowest accuracy scores. The test setup, in the post’s words, was distributed benchmarking on EC2 with a tool it calls Harbor. EC2 is Amazon’s rented computers. The post says a follow-up paper from its researchers is still to come. That paper is not on this page, and the six benchmarks are not named here. File the 28 percent, the nearly equal scores, and the Deepseek caveat as the company’s.

A second company number, still not a desk rerun. With a model the post calls Fable 5, Strands harness cost 77 percent less than Claude Code and scored higher on a test the post calls Terminal Bench 2.1. A terminal is the text window on a computer. Terminal Bench is a test of whether an agent can finish tasks there. The post also says two other open-source harnesses did similarly well on cost and accuracy against Claude Code. It does not name those two. File the 77 percent and the higher Terminal Bench score as the company’s claim about Fable 5. This desk did not run that test.

What the default agent includes. The post says one call, named create_harness, returns an agent that runs on a current model across Amazon Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM. Bedrock is Amazon’s service for calling models, including models from other companies. Ollama is a way to run a model on your own computer. LiteLLM is a switchboard that speaks several companies’ model formats. Out of the box the agent has shell, file, and web tools. A shell tool runs commands. The file tools are named read, write, and edit. A web tool can look things up online. The agent manages its own context window, which is how much text the model can hold at once. Tool results over about 1,500 tokens get truncated, which means cut down. Summarization, which the post also calls compaction, starts when that window is above 85 percent full. If the window still overflows, a recovery step runs inside the loop. The agent keeps long-term memory across runs and can resume an earlier conversation when you give it a session ID, an identifier for that conversation. It can hand an open-ended subtask to a built-in helper agent. It loads skills if you have any. File that default list as the post’s.

Where it can run. The post says Strands harness can be deployed on any provider that offers a Linux container. A container is a packaged program that runs the same way on different hosts. Linux is the operating system those hosts are using. The examples named are Modal, Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS, and Amazon Bedrock AgentCore. Locally, or any cloud, is the post’s own frame. This is not a product the page locks to Amazon. File that list as the post’s. This desk did not deploy it.

Do not read the cost lines as an outside ranking. The 28 percent figure and the 77 percent figure are the project’s own benchmark claims. On the same page, Deepseek Harness was more token-efficient and less accurate. The six benchmarks are unnamed. The follow-up paper is not published here. A harness you can install is not a score this desk measured, and it is not a claim that every other harness costs more.

Plain English for the rest of the card: harness = the assembled agent around a model, including tools, memory, and the rules for how long a conversation can get. Apache 2.0 = an open-source license that lets you use and ship the code if you keep the notices. general-purpose agent = an agent for many jobs, which this page sets apart from a coding agent. token = a small chunk of text you pay for. context window = how much text the model can hold at once. compaction = the post’s word for summarizing that window once it is above 85 percent full. session ID = the tag that resumes an earlier conversation. Bedrock, Anthropic, OpenAI, Google, Ollama, and LiteLLM = the model providers the post names. container = a packaged program you can run on different clouds. This filing is the 21 Sep release. It is not a benchmark this desk ran.

PRIMARY here: the Strands Agents post “Introducing Strands harness,” dated September 21, 2026, on strandsagents.com — Tier A PRIMARY, the project’s own record. The public Python package page for strands-harness is the same-day package record, not a second newsroom. The Apache 2.0 release, the general-purpose framing, the one-line Python and TypeScript setup, the three install commands, the /export step, the 28 percent claim on the same Claude or GPT models across six benchmarks, the nearly equal scores against Claude Code, Codex, and other popular harnesses, the Deepseek Harness caveat, the EC2 and Harbor setup line, the promised follow-up paper, the Fable 5 77 percent claim and the higher Terminal Bench 2.1 score, the two unnamed open-source harnesses, the default tools, the 1,500-token truncation, the compaction above 85 percent, the overflow recovery, the memory and session ID, the helper agent, the skills line, the Bedrock, Anthropic, OpenAI, Google, Ollama, and LiteLLM providers, and the any-Linux-container deploy list are the post’s. The same-day upload stamps are the package registries’. NOT claimed: an independent benchmark rerun, the names of the six benchmarks, the names of the two other open-source harnesses, that Deepseek Harness costs more, that the product runs only on Amazon, that this desk installed the package or deployed a container, a stock tip, or investment advice. Distinct from the already-filed minimax-code, zai-zcode-open-source, and azure-cosmos-spec-kit.