NederlandsAmazon has AI agents build an entire testing program for its cloud
According to Amazon, thousands of AI agents wrote all the code of AWS Emulator, a free program that imitates Amazon's cloud services on a developer's own computer. People designed the way of working and the checks, but did not review the code itself.
Amazon Web Services (AWS), the cloud division of Amazon, released AWS Emulator . It is a free program that imitates AWS's online services on a programmer's own computer. Anyone building software that has to run in the AWS cloud, on the computers Amazon rents out over the internet, can now try it out on their own computer first, without paying for it.
The striking part is how the program was made. According to Amazon, thousands of AI agents wrote all the code. Testing, finding bugs and making improvements are also done by agents, without a human reviewing the code. Amazon engineer Matthew Goldberg writes that his team did not write or review any code itself.
Three crews of agents
According to Amazon, the agents work in three crews. The first writes tests: small checks that verify whether the imitated service gives the same answer as the real AWS service. The second runs those tests again and again and reports every difference as a problem. The third picks up such a problem and solves it.
Before an improvement is added, yet other agents check whether the code is tidy, whether it is safe and whether it breaks anything else. A separate security agent checks every change twice: during the work and once more just before a new version is released.
People did design this way of working and the checks, and according to Goldberg that still takes a good deal of technical work. It started earlier in 2026 with a weekend experiment. He asked Kiro, Amazon's AI coding assistant, to build a first imitated service, and after fifteen minutes there was a usable first version. Amazon does not say which AI models the agents now use.
According to programmer Simon Willison, 29,684 changes were added to the code in the first three weeks of the project. Anyone can view and use the code of AWS Emulator.
Not equally good everywhere yet
AWS uses a public set of tests to measure how well the emulator imitates the real services, and publishes how many tests pass for each service. That still differs from service to service. When new tests are added, the score can drop, even if the emulator gets better.
There are also limits. The emulator knows only one account and does not check who is allowed to do what in the same way as AWS. A request that the real AWS would refuse may therefore simply succeed in the emulator. For that reason, AWS advises against using the emulator to test whether software is secure, and recommends testing everything on AWS itself before real use.
What does this mean for the road to the singularity?
A large, usable program built and maintained by AI agents shows that AI can do more than write separate bits of code: it can make and maintain a whole product. That works here because the agents keep testing their work against the real AWS services and because separate agents check every change. Whether the same approach also works for software without such a real example to compare against is the next question.
What is the technological singularity? Read the explanation →
