Amazon says nobody on the team wrote or reviewed a single line of AWS Emulator. That claim is easy to roll your eyes at. But the way they got there is more useful than the headline.
AWS announced a local emulator of AWS services that runs on your machine, for building and testing apps without hitting the real cloud. It’s released on GitHub under a CC0 license, and agents keep updating it.
A weekend experiment that stuck
The project began after Anthropic showed agents could autonomously build a C compiler. The author’s reasoning was simple: give agents a spec and a reference implementation and they can match it. AWS has both. The API Models are the spec, and the live cloud is the reference. He prompted Kiro, Amazon’s coding agent, to write CLI-based tests, make them pass against real AWS, then build an in-memory version that passes the same tests. Fifteen minutes later, a passable SQS appeared. By Monday, his team was using a rough version of the rest.
Getting from there to something customers can trust took far longer. That’s the honest part of the story.
Three agent teams and a merge gate
The architecture splits work into bounded jobs:
- Test team: writes tens of thousands of fidelity tests from the API Models and verifies each against real AWS first.
- Finder team: replays those tests against the emulator and files grouped issues for failures.
- Fixer team: takes one issue each, edits the code, and submits it.
Fixes then hit a merge gate. It checks compilation, coding standards, security, and whether fidelity regressed. Fail, and the change goes back for rework. Hundreds of agents run at once, and thousands of issues get fixed daily.
How it compares
LocalStack is the obvious incumbent here, and Moto is the open-source favorite for Python mocking. Both are built and maintained by humans. AWS’s version covers all 438 API Models in principle, though gaps remain and the backlog is public in the repo. That transparency is a nice touch.
The release path is also telling. Code moves through AWS developer tooling and lands on GitHub only at the end.
Does it matter?
Yes, but for the method more than the product. Test suites that run against a real reference system give agents an unambiguous target. Few domains have that. Anything with a strong spec and a live oracle, like databases, protocols, or SDKs, could follow the same pattern.
The caveat is that the agents once built their own Python interpreter for Lambda before being told to reuse existing libraries. Autonomy without guidance gets weird. So watch the guardrails, not the agent count. That’s the part worth copying.
