Amazon released AWS Emulator on Wednesday: a free, locally hosted software suite that replicates several hundred AWS services on your own machine. The remarkable part is in the blog post. Matt Goldberg, a senior principal engineer at AWS, writes that the entire service was designed and written exclusively by AI agents, with not a single line of code written or reviewed by a human engineer.
The problem Amazon is trying to solve
The emulator answers a bottleneck that coding agents created themselves. Tools like AWS Kiro and Claude Code produce code faster than teams can verify it. Verifying that code requires real infrastructure, and doing it in the real AWS cloud costs both time and money. The third-party emulators that already exist cover a fraction of the AWS catalogue and fall behind the services as those evolve.
Goldberg’s route: he took AWS’ public Smithy API models as the formal specification and the running cloud as the reference implementation. The nudge came from research showing that agents can build a C compiler from scratch. If that works, he reasoned, then any AWS service can be rebuilt as long as a spec and a reference are available.
A twenty-year-old story as the argument
Werner Vogels handed his blog over to Swami Sivasubramanian, now vice president of agentic AI, for the announcement. Sivasubramanian tells the story of “Alf”, a Paxos-based distributed lock service he built with Al Vermeulen about twenty years ago, which still sits underneath S3 and DynamoDB.
Back then the only way to test Alf was to simulate its entire environment inside a hand-built Java VM, which took months. They had to inject faults themselves and isolate concurrency bugs by hand. At the scale AWS runs today, hand-crafting that simulation is no longer feasible, and that is the real point of this launch: the emulator gives agents the same kind of test bench that work depended on.
Why the test bench matters more here than the author
Agents wrote the code, and that is the headline, which is how it ran on Hacker News too. What interests me more is what they wrote it against. A machine-readable specification, a running reference you can query, and a test cycle that answers in seconds rather than minutes. Those three conditions were in place at AWS and are missing from most projects.
Take away only that agents now write software and you have half of it. The other half is that this half repeats only where somebody maintained the specification first. The project is public on GitHub, at nearly 2,500 stars a day in, with v1.0.322 as the latest release.