AWS has released a free, open source emulator that it says thousands of autonomous AI agents wrote with no human code review. The AWS Emulator lets developers run supported AWS services on their own computers or in continuous integration jobs. Applications connect through a local endpoint using the same APIs, SDKs and command line tools. It needs no AWS account for local testing, and the code sits on GitHub under a CC0 license.
The project began as a small experiment. Matthew Goldberg, a senior principal engineer at AWS, asked Kiro to build an in memory version of Amazon SQS, and the first attempt took about 15 minutes. From there, it grew into three groups of agents. One writes tests that compare behavior with real AWS services, one finds differences and logs them as issues, and one fixes them. Goldberg says hundreds of agents work at once and clear thousands of issues a day, while AWS engineers oversee the system without writing or reviewing the code. Each change must pass automated checks for compilation, standards, security and compatibility, and a separate security agent runs adversarial reviews.
The accuracy claims come with caveats. AWS publishes a pass rate for each service from its fidelity tests, and the repository says AWS checked each test against at least two sandboxed AWS accounts. Still, the repository itself says behavior outside those tests is unverified, and operations the emulator has not built return an HTTP 501 error. It matches production most closely for creating, describing, updating and deleting resources, while storing and retrieving data varies by service.
Security gaps deserve attention too. The emulator does not evaluate IAM policies the way AWS does, so a request that AWS would reject for missing permissions could succeed locally. The emulator does not sandbox code, so anything it runs can reach local files with the developer’s permissions. Devin Pratt of IDC says the first release skips throttling, outages and regional differences, and Advait Patel of Broadcom says timeouts and partial failures still need real infrastructure tests.
AWS tells developers to run only trusted code, keep real credentials and production data out of the test setup, and try security settings and full applications in AWS before launch. The published scores will show whether the agents close those gaps.

