This week: Amazon released AWS Emulator, a free tool for testing cloud apps on your own computer, and says AI agents wrote every line of its code.

Elsewhere: How three teams of agents grade each other's homework, what the emulator can't do yet, and three fresh AI drops worth a skim.

The New Prime Signal

Most announcements about AI writing code come with an asterisk the size of a beach ball: a human cleaned it up, a human reviewed it, a human quietly rewrote the hard parts. This one tries to skip the asterisk. Amazon released AWS Emulator, a program that pretends to be Amazon Web Services on your own machine so you can test cloud apps without touching the real cloud. According to the announcement post by Amazon's Matthew Goldberg, the emulator was written entirely by autonomous agents, and his team hasn't written or reviewed any of the code. Thousands of agents, one product, zero human pull requests. Naturally, the first question is whether it actually works.

What shipped: AWS Emulator is a local stand-in for AWS APIs. Per the AWS Emulator User Guide, you don't need an AWS account, there's no extra service fee, and your requests go to the emulator instead of AWS. On Linux you install it from an RPM or Debian package; on macOS and Windows you run it in Docker; and Rust developers can run it inside their test process. The emulator keeps whatever you create, so if you make a storage bucket and list your buckets later, it's there. If you try something AWS would reject, like a database table name that's too short, the docs say you get the same error AWS would give you. The source code is on GitHub under a CC0 license, which is about as close to "take it, it's yours" as licenses get, and the agents' fidelity tests are published too.

Why it matters: Testing against the real cloud costs money, needs credentials, and can be slow. A local copy you can spin up in a test and throw away afterward is the kind of unglamorous tool developers quietly love. The bigger story is how it was made. Goldberg writes that the idea clicked after Anthropic researchers showed agents could build a C compiler on their own. AWS had the two ingredients that made that work: a written spec (its public API models) and a reference to check against (the actual AWS cloud). If agents can build something useful whenever both of those exist, a lot of software is suddenly on the table.

The fine print: The docs are refreshingly blunt. AWS Emulator is for building and testing, not production. It doesn't evaluate access policies the way AWS does, so AWS says not to use it to test security. It keeps everything in memory and deletes it all when it stops, it hasn't been tested for long-running or load testing, and it checks its behavior against one AWS region. Per the fidelity page, it matches AWS most closely for the APIs that create and manage resources, while the day-to-day data side varies by service, and anything not built yet returns a "not implemented" error. It also sends anonymous usage data by default, which you can switch off.

The New Prime Signal take: "No human reviewed the code" will make some engineers spill their coffee. Fair. But Amazon isn't asking for blind trust. It's betting on tests that were checked against the real thing, and it's publishing the tests, the backlog, and the code so anyone can poke holes. That's a more honest pitch than most "AI wrote this" headlines.

This started as a weekend experiment. Goldberg says he gave Kiro, Amazon's coding agent, a three-step prompt: write tests based on the AWS command line tool, keep fixing those tests until they pass against real AWS, then build an in-memory version that passes the same tests. Kiro came back after 15 minutes with a passable first version of Amazon SQS, AWS's message queue. He refined the prompt and had Kiro build "the rest of AWS" over the weekend. It had obvious gaps, he writes, but his team started using it that Monday. Turning that into a real product took a lot more than a weekend.

How it works: According to Goldberg's post, the work is split among three teams of agents, because they do better on bounded tasks.

The best detail: an earlier version of the agents decided to build their own Python interpreter to support AWS Lambda functions. They've since been told to reuse existing libraries, which Goldberg calls obvious to "someone who doesn't need a coloring book."

Try it: Start with the user guide, which walks through Linux, Docker, and AI-coding-agent setups. If you want to watch the machine work, the releases page lists each batch of agent fixes, and the GitHub repo keeps every issue in two folders: things-that-now-work and things-that-do-not-work. Building it from source is a bigger ask. The README says you'll want at least 100 GB of free disk space. Most people should grab a prebuilt package instead.