Amazon offers a tool on GitHub whose AI agents wrote 100% of the code. Distributed under CC0 license and continually updated, the project started with an experiment carried out over a weekend.
The tool replicates APIs without covering all of AWS
The name emulator doesn’t exactly match how it works: this application replicates AWS API endpoints locally, but not all cloud functions. Developers can thus test their software before full deployment and find problems before moving into production. This local check aims to prepare for code execution on AWS.
AWS indicates that these tests can take place without requesting its services, either on a personal computer or on a server dedicated to development or in a continuous integration process. For Mac and PC environments, execution goes through Docker. Under Linux, RPM or Debian packages are offered depending on the distribution.
Goldberg distinguishes rapid prototype and robust product
Development took place without code review by humans on the AWS team. Matthew Goldberg specifies, however, that the transformation of the prototype into a robust product took more than a weekend. He explains in particular that the team had to put in place safeguards so that the agents respected the limits it had defined.
Deciphered AI dev tools, agents and new features, every evening in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Goldberg warns against a reading of these results that would make AI a system already capable of building anything. In his eyes, offering customers a product deemed sufficiently reliable always requires significant engineering work.
Kiro produced a working version in 15 minutes
At the origin of the experiment was a demonstration by Anthropic: agents built a C compiler independently. The AWS team then provided Kiro, its autonomous agent, with a precise specification through a simple prompt. This instruction, organized in three stages, was to test a hypothesis of Matthew Goldberg, senior principal engineer.
Kiro had to first write tests with the AWS Command Line Interface, then rework them until they were validated on AWS itself. He then had to build an implementation that would pass these tests. A working version was obtained in 15 minutes, to the surprise of the team. By reworking the prompt, Goldberg had a version that his colleagues could use first thing Monday morning.

