AWS Emulator runs 200 Amazon services offline, and no engineer read its code

AWS Emulator, Amazon‘s new developer tool, puts a working stand-in for its cloud on an ordinary computer, so software that normally talks to Amazon’s servers can be tested offline, for free and without an account. The surprise is who built it. Amazon says AI agents wrote every line, and the engineers on the project have not written or reviewed any of the code.

For people who build software, that changes a slow and expensive habit. Testing an app against the real cloud means waiting on servers and paying for resources that exist only to be thrown away. Coding agents such as Amazon’s Kiro and Anthropic‘s Claude Code make the problem worse, because they check their work by running code over and over, and every round trip to the cloud adds delay and cost. The emulator lets that loop run on a laptop instead.

The scale is what stands out. The tool covers more than 200 AWS services, generated from over 400 of Amazon’s public API descriptions, and it is written in Rust. According to Matt Goldberg, the AWS senior principal engineer behind the project, it began as a weekend experiment. He gave Kiro a single prompt, got a passable copy of Amazon’s message-queue service in about 15 minutes, and by Monday had a rough version of the rest of AWS that his team started using to test its own software.

What keeps the code honest is not a human reviewer but a chain of automated checks. Amazon describes three teams of agents running on its Bedrock AgentCore platform. One writes “fidelity tests”, real scenarios such as creating a queue, sending a message and deleting the queue, and verifies each one against the live cloud. A second replays those tests against the emulator and files every failure as an issue. A third fixes them, one agent per issue, and nothing merges until other agents, several of them dedicated to security, have attacked the change. Goldberg says hundreds of agents now fix thousands of issues a day.

His own explanation is disarmingly low-tech. Goldberg compares his agents to himself in kindergarten, too young to hold a crayon properly and blamed for spoiling a class drawing of dinosaurs, and concludes that agents need coloring books: a clearly outlined area to fill in. Without that structure, an earlier generation of the agents decided to write their own Python interpreter to run Lambda functions, a detour the team has since steered them away from.

The limits are real, and Amazon publishes many of them itself. The project’s documentation tells developers not to use the emulator for production workloads or to test security, and the repository keeps a public folder of behaviors that still don’t match real AWS, named “things-that-do-not-work”. IDC analyst Devin Pratt told InfoWorld that the tool does not yet simulate throttling, outages or regional differences in which services are available, and that it does not sandbox the code it runs. Building it from source needs at least 100 GB of free disk space, and telemetry that reports usage to Amazon is on by default, although it can be switched off.

It also arrives in a market with an established player. LocalStack, the best-known AWS emulator, this year put its free version behind an account requirement and limited it to noncommercial use, as GeekWire reported. Amazon’s tool is released under CC0, a public-domain dedication, with the full commit history and every agent-filed issue on GitHub, so anyone can inspect it, fork it or build on it.

The release is a statement as much as a product. Swami Sivasubramanian, AWS vice president of agentic AI, wrote on the blog of Amazon CTO Werner Vogels that when agents write the code, the engineer’s job becomes building the specs, tests and merge gates that decide when the work is correct. He described supervising Kiro for four nights as it tried to rebuild a replication engine he had built by hand about twenty years earlier, a job that took about two hours once the agent had a spec and a test environment.

AWS Emulator went live on November 11 as a free download worldwide, with packages for Linux, a Docker route for Windows and macOS, and an embedded mode for Rust test suites. Amazon says it will share more about the project at its re:Invent conference in a few weeks.

Whether teams will trust code that no person has read is the question Amazon has chosen to put in public. The folder of things that do not work is where most developers will start looking.

Tags: amazon, Open Source, AI coding agents, Amazon Web Services, AWS Emulator, Kiro