AWS Emulator: agents wrote it, real AWS grades it, nobody reviewed it
AWS just shipped a developer tool and says no engineer wrote or reviewed a single line of it. AWS Emulator, released Wednesday, runs a local copy of more than 200 AWS services on your own machine, built from over 400 public API models. The AWS CLI, the SDKs and coding agents talk to it the way they talk to the cloud, with no network connection and no bill. It is free, released under CC0 with the full commit history on GitHub: https://github.com/aws/aws-emulator, launch post at https://aws.amazon.com/blogs/aws/test-your-applications-against-aws-locally-with-aws-emulator/.
The build story is the real news. "The emulator was written entirely by autonomous agents. Our team hasn't written or reviewed any of the code," writes Matt Goldberg, the senior principal engineer behind it (https://www.aboutamazon.com/news/aws/how-autonomous-ai-agents-built-aws-emulator). The agents run on Bedrock AgentCore in three teams. One writes fidelity tests, small real scenarios like create a queue, send a message, delete the queue, and checks each one against real AWS. A second runs those tests against the emulator and files issues for every failure. A third fixes them, one agent per issue. Every change goes through a gate of other agents checking standards and security. Goldberg says hundreds of agents now fix thousands of issues a day. It started as a weekend experiment: Kiro produced a passable SQS in 15 minutes from a three-step prompt.
Notice where the trust sits. Not in the code, which nobody read, but in the oracle. Real AWS is the answer key, and the agents only get to declare victory when their copy behaves like the original. That is the cleanest version yet of a pattern this year keeps repeating: the verifier is the product. Swami Sivasubramanian said it plainly in a guest post on Werner Vogels' blog: the engineer's job is now "the specs, tests, and merge gates that decide when that work is correct." Sivasubramanian's own anecdote is the best argument for it: four nights babysitting Kiro on a replication engine, then about two hours once the agent had a spec and a way to test itself.
Two cautions. "Nobody reviewed the code" is AWS's selling point, not an independent audit. And IDC analyst Devin Pratt told InfoWorld the emulator does not yet simulate throttling, outages or regional differences, and it does not sandbox the code it runs. Treat it as a fast test loop, not a safe box. The competitive angle is sharp too: LocalStack, the best-known alternative, put its free tier behind an account and limited it to noncommercial use this year. AWS just answered with a public-domain clone that its own agents keep current. For anyone whose coding agent burns time and money re-running tests in the cloud, the inner loop just moved to localhost.
← Back to all articles
The build story is the real news. "The emulator was written entirely by autonomous agents. Our team hasn't written or reviewed any of the code," writes Matt Goldberg, the senior principal engineer behind it (https://www.aboutamazon.com/news/aws/how-autonomous-ai-agents-built-aws-emulator). The agents run on Bedrock AgentCore in three teams. One writes fidelity tests, small real scenarios like create a queue, send a message, delete the queue, and checks each one against real AWS. A second runs those tests against the emulator and files issues for every failure. A third fixes them, one agent per issue. Every change goes through a gate of other agents checking standards and security. Goldberg says hundreds of agents now fix thousands of issues a day. It started as a weekend experiment: Kiro produced a passable SQS in 15 minutes from a three-step prompt.
Notice where the trust sits. Not in the code, which nobody read, but in the oracle. Real AWS is the answer key, and the agents only get to declare victory when their copy behaves like the original. That is the cleanest version yet of a pattern this year keeps repeating: the verifier is the product. Swami Sivasubramanian said it plainly in a guest post on Werner Vogels' blog: the engineer's job is now "the specs, tests, and merge gates that decide when that work is correct." Sivasubramanian's own anecdote is the best argument for it: four nights babysitting Kiro on a replication engine, then about two hours once the agent had a spec and a way to test itself.
Two cautions. "Nobody reviewed the code" is AWS's selling point, not an independent audit. And IDC analyst Devin Pratt told InfoWorld the emulator does not yet simulate throttling, outages or regional differences, and it does not sandbox the code it runs. Treat it as a fast test loop, not a safe box. The competitive angle is sharp too: LocalStack, the best-known alternative, put its free tier behind an account and limited it to noncommercial use this year. AWS just answered with a public-domain clone that its own agents keep current. For anyone whose coding agent burns time and money re-running tests in the cloud, the inner loop just moved to localhost.
Comments