Another week, another AI agent announcement. But this time it is Amazon doing the humble-pie move: instead of selling you the harness, it is giving it away. AWS open-sourced Strands Harness on September 21 under an Apache 2.0 license, and the internet promptly started arguing about its math.
The pitch, in one line: a ready-to-use, general-purpose agent framework that runs locally or in any cloud, sips fewer tokens than the big-name coding agents, and comes with benchmarks that AWS graded entirely by itself. Which, as any student will tell you, is the best kind of grading.
What Is Strands Harness, Anyway?
Strip away the marketing and Strands Harness is the scaffolding around a model: the loop, the tools, the context window, the memory, the recovery. Most developers have built a working AI agent prototype on a laptop with tools like Claude Code or Codex "because they just worked" on the local machine. Strands Harness is AWS's answer to what happens when that prototype has to grow up and move to a cloud.
It ships with read, write and edit tools, shell access, web search, a helper agent for open-ended subtasks, and long-term memory that resumes conversations across runs via session IDs. Developers can install it with a single pip or npm command, point it at any frontier model from Amazon Bedrock, Anthropic, OpenAI or Google, or let it run completely offline on local models through Ollama and LiteLLM.
Oh, and it runs on anyone's cloud. Amazon, Google, Azure, Cloudflare, Modal, your neighbor's Linux container. Apache 2.0 means AWS cannot even stop you from running its harness on a competitor's infrastructure, which is either generous or the most passive-aggressive multi-cloud strategy ever devised.
The Number Salad
Here is where things get fun. AWS's own team benchmarked Strands Harness on Harbor, the evaluation framework from the Terminal-Bench creators, and the headline numbers are a buffet:
- 28% fewer tokens than rival harnesses at comparable accuracy across Claude or GPT comparisons
- 77% less cost than Claude Code on the same tasks in one Fable 5 test, with a higher score to boot
- 26% more efficient than agents built on other frameworks using the same underlying model
- 45% cheaper than Claude Code and Codex, per one outlet's reading of the docs
- Oh-my-pi matched its 69.7 accuracy at 54% higher cost, and DeepSeek Harness was somehow even cheaper, at 14% under Strands, while scoring lower
Read that list again. AWS is 28% cheaper, 45% cheaper and 77% cheaper at the same time. The secret is that these are different comparisons, which is exactly how this genre of press release works.
The Register, ever the class clown, noted the obvious: Strands "only raced coding agents and marked its own homework." The benchmarks were run by the Strands Agents team itself, on EC2, using models of its choosing. Nobody is saying the numbers are fake. Everybody is saying they are conveniently self-reported.
Why Everyone Is Suddenly Cheaper Than Everyone Else
Underneath the marketing chaos there is a real story. The cost of running agents is now the battleground, and the harness around the model matters more than the model does. Independent research backs this up: the HarnessTax study compared Claude Code, Codex CLI and Pi across seven models and found that harness choice barely moved success rates, while the same model could cost up to five times more depending on the wrapper.
The real secret sauce is context management, squeezed into three blunt rules:
- Truncate tool results that exceed about 1,500 tokens
- Start context compaction when the window hits 85% capacity
- Recover and rebuild context when an overflow happens anyway
Toss in aggressive prompt caching and you get the token savings that power the entire marketing campaign. The boring plumbing is the product.
Then there is the Strands CLI, which lets you describe an agent in plain English and exports the whole thing to Python or TypeScript with a single /export command. Build an AI agent like you are ordering a burrito: tell it what you want, watch it assemble, then take the receipt and run. AWS engineers built the CLI on top of the harness itself, the developer equivalent of eating your own dog food, if the dog food were open source and free.
So what is the takeaway from the open-source AI week? The agent harness market has officially become a price war where every vendor is cheaper than every other vendor, by their own accounting. For developers, that is genuinely good news: Apache 2.0 code, local-first defaults, and a framework that follows your prototype from laptop to production without being rewritten.
Just maybe ask for a second opinion on the homework grades before you bet your token budget on them.
Comments