Amazon says AI agents wrote and reviewed AWS Emulator code

Amazon released AWS Emulator on November 11 to let developers test cloud applications locally by simulating AWS services. Amazon says autonomous AI agents wrote and reviewed the tool's implementation. Engineer Matthew Goldberg says in Amazon's engineering account that Goldberg's team neither wrote nor reviewed the emulator code. Goldberg describes substantial engineering work to prepare the experiment for release, including safeguards and checks governing which changes can ship.

Amazon describes three agent teams with separate responsibilities. One team maintains fidelity tests. Another team identifies differences between the emulator and real AWS. The third team resolves individual issues. Amazon says proposed changes must pass compilation, coding-standard, security and regression checks before merging. The testing agents use public AWS API models and validate scenarios against real AWS. The public evaluation suite states that every fidelity test underwent verification against at least two sandboxed AWS accounts.

AWS uses the term "fidelity" for agreement between the emulator and AWS cloud services. AWS's fidelity documentation says a passing workflow must match AWS on every call, including status codes, resulting resources, side effects and errors. A test fails if the emulator reports success when AWS reports an error. The evaluation repository states that behavior without a fidelity test remains unverified. A successful scenario demonstrates agreement only for that scenario, not compatibility with every way an application might use AWS.

AWS's security guidance warns developers not to use the emulator to assess whether an application is secure. The emulator does not evaluate AWS KMS key policies or Amazon S3 bucket policies. Requests that AWS rejects may succeed in the local emulator. Agent review of implementation security therefore does not validate permissions in a developer's application. The emulator supports development and testing rather than production workloads. AWS instructs developers to test the entire application and security settings on real AWS before production.

For developers adopting software written by agents, the release makes acceptance checks as relevant as authorship. Those checks include verified behaviors, gaps in test coverage and areas that still require testing against the real service.