AWS Ships a Tool No Human Wrote or Reviewed: Hundreds of AI Agents Wrote the Tests, Found the Bugs and Fixed the Code

Amazon Web Services on Nov. 11 released AWS Emulator, a free developer tool that stands in for AWS cloud services on your own computer. The tool itself isn't the headline. Its origin is: AWS says every line of its code was written by AI agents, and the engineers on the team neither wrote nor reviewed any of it.

What it does

Lots of companies run the back end of their websites and apps on cloud platforms like AWS, using it to store files, run databases and pass messages between systems. Developers want to test every code change, but testing against the real cloud each time means waiting on deployments, paying for resources and risking live data.

AWS Emulator is a local stand-in. It imitates AWS services, so a developer's program, command-line tools and AI coding assistants send requests as usual, but those requests go to their own machine instead of the real cloud. According to GeekWire, it's built from more than 400 API models covering more than 200 AWS services and works without a network connection. The AWS News Blog says it carries no service fee and its source is on GitHub. A post by AWS senior principal engineer Matt Goldberg says the code is released under CC0, a public dedication that effectively waives copyright so anyone can use it.

A few terms

AI agent: an AI that doesn't just answer questions but carries out a chain of tasks on its own, such as writing code, running it, reading the results and fixing it.

API model: a spec that defines what requests each cloud service accepts and what it should return. AWS treats these as the outlines in a coloring book for its agents.

Fidelity test: a small real-world scenario, such as "create a message queue, send a message, delete the queue." It's run once against real AWS to record the correct behavior, and then the emulator has to respond exactly the same way.

How the AI built it

Goldberg writes that the project began as a weekend experiment. Inspired by Anthropic's demonstration earlier this year in which a team of Claude agents built a C compiler, he gave AWS's own coding agent, Kiro, a three-step prompt: write tests based on the command-line tools, iterate until they pass against real AWS, then build an in-memory version that passes the same tests. Fifteen minutes later, Kiro returned a passable first version of the Amazon SQS message-queue service. By the end of the weekend it had built a rough version of the rest of AWS, and his team started using it to test their software that Monday.

Turning that into a product took a strict division of labor and a series of checkpoints. As Goldberg describes it, the agents work in three teams. One writes fidelity tests from the API specs and verifies each against real AWS. A second replays those tests to find where the emulator differs from AWS and files issues. In the third, each agent fixes a single issue. Every change must pass an automated gate: code that won't compile is blocked, and other agents check coding standards, security and whether the change broke anything that used to work. Any failure sends the change back for rework. The AWS News Blog adds that an independent security agent reviews changes adversarially, once during the fix and again in the release pipeline. Goldberg says hundreds of agents now work on the codebase at once, fixing thousands of issues a day.

One telling detail: in an earlier version, the agents decided to write their own Python interpreter to support AWS Lambda functions. The team has since told them to reuse existing libraries.

Where it falls short

The AWS News Blog is explicit about the limits. The emulator doesn't evaluate access-control (IAM) policies the way AWS does, so a request the real cloud would deny might succeed locally, which means it can't be used to test whether an app is secure. Each emulator acts as a single account, with no cross-account or cross-Region behavior. Fidelity varies by service, and with each release AWS publishes how many fidelity tests each service passes, including the ones that fail. When a feature isn't implemented yet, the emulator returns an error so tests fail right away instead of quietly passing.

Why it matters

"A human has to review AI-written code" has been close to an industry rule. Here one of the world's biggest cloud providers is shipping a production tool and advertising that no human reviewed the code. It works because this job happens to come with a clear spec (the API models) and a reference to check against at any time (the real AWS). In a guest post on CTO Werner Vogels' blog, Swami Sivasubramanian, AWS vice president of agentic AI, wrote that when agents write the code and check their own work, the engineer's job is to give them room and to build the specs, tests and merge gates that decide when that work is correct.

That points to a shift in what software engineers do: less writing code by hand, more writing rules, setting the exams and acting as referee. It also raises open questions. Who is accountable when code nobody read goes wrong? And does the approach carry over to software that has no reference answer to check against? Nobody has settled those yet.

What it means for you

You won't use this tool directly, but many of the apps on your phone run on the cloud. Faster, cheaper testing for developers should, in principle, mean new features arriving sooner with fewer bugs. The bigger point is the signal: AI writing, testing and fixing software on its own has moved from demos into a major company's shipping product. More of the software you use may come to include code no human has read, and how trustworthy it is will depend on how strict the rules people set for the AI are.

Original sources published: guest post on CTO Werner Vogels' blog, Nov. 11, 2026, 22:30 Beijing time (UTC+8); About Amazon post by Matt Goldberg, Nov. 12, 2026, 01:25 Beijing time; GeekWire, Nov. 12, 2026, 02:51 Beijing time; AWS News Blog, Nov. 12, 2026, 03:00 Beijing time