Timeline

Amazon says autonomous agents wrote its new AWS emulator

Amazon says its team neither wrote nor reviewed the code; three teams of agents write tests, file issues and merge fixes through a gate, and the source is released under CC0.

  • Labs & people
  • Ideas & essays
  • Notable

Amazon released AWS Emulator, a tool that runs AWS services on a developer’s own machine or in a CI environment, and said in an Amazon News post that the code was “written entirely by autonomous agents”. “Our team hasn’t written or reviewed any of the code,” its author wrote. AWS’s announcement says calls from the AWS SDK and CLI go to the emulator, which keeps resource state in memory, charges no service fee, and returns “501 Not Implemented” for operations it does not yet support. The source is published under the CC0 public-domain dedication.

The post describes a weekend experiment that became a process. Its author said the idea came from Anthropic’s demonstration that agents could build a C compiler, and reasoned that AWS has what such a project needs: published API models as a specification and the real cloud as a reference implementation. Given a simple prompt to write tests from the AWS command-line interface, iterate until they passed against real AWS, and build an in-memory version that passed the same tests, Amazon’s coding agent Kiro produced a passable first version of the SQS queue service in 15 minutes, and the rest of the emulator, with gaps, over a weekend.

Turning that into a product, Amazon said, took guardrails. Three teams of agents divide the work: one writes fidelity tests from the API models and checks each against real AWS, one replays those tests against the emulator and files issues for failures, and one fixes one issue per agent. Every fix passes a merge gate that checks the build, coding standards, security and that fidelity has not regressed. Amazon said it has tens of thousands of fidelity tests, and sees thousands of issues fixed a day by hundreds of agents working at once. The post also recounts an earlier version in which the agents wrote their own Python interpreter to support Lambda functions, before the team told them to reuse existing libraries.

These are Amazon’s claims about its own process. Neither page gives a figure for how closely the emulator matches AWS: the announcement links a fidelity report, which was not reviewed here, and the post points to the public repository, where Amazon says every filed issue can be read. The post also cautions against the idea that “AI can build anything”, saying a product customers can trust still took serious engineering effort.

The release is a worked example of the pattern Anthropic’s compiler showed earlier in the year, agents building a large program against a spec and a test suite, now shipped as a vendor product. Amazon said it is the first of a series of posts on how it architected the workflow and measures fidelity.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.