Amazon Web Services (AWS) launched this Wednesday (11) the AWS Emulator, a development tool whose code was written entirely by autonomous artificial intelligence agents, without human review of each change. According to the company, thousands of agents took part in creating the software, which allows developing and testing cloud applications directly on the computer.
Matthew Goldberg, senior principal engineer at AWS, said the human team did not take part in the programming or the review of the code produced. “Our team did not write or review any part of the code”, he stated in a technical account published by Amazon.
The emulator reproduces locally the behavior of AWS APIs, allowing developers to run tests without having to provision resources in the cloud. The tool is available for download, with no additional charge for using the emulator.
How AWS organizes agents to test and fix code
The development was divided into three teams of autonomous agents, each responsible for a specific stage.
The first creates tests based on the API specifications and verifies how they work on the real AWS infrastructure. The second runs these same tests on the emulator to identify differences in behavior. The third receives the problems found and modifies the code to fix them.
Before integrating any change, the agents must pass automated build, quality, security and compatibility checks. Changes that do not pass these verifications return for correction.
Goldberg says the system already gathers tens of thousands of fidelity tests and can resolve thousands of problems daily, with hundreds of agents working simultaneously.
Human supervision, in this model, focuses on defining the specifications, the development rules and the validation mechanisms, instead of the individual review of lines of code.
AWS made available the source code and the tests on GitHub, with the project published under the CC0 license, which allows broad reuse of the material.
Despite the autonomy described by the company, the emulator has compatibility limitations with the real services. The official documentation also clarifies that the tool is intended for development and testing, not for running applications in production or for security validation.
