In an announcement published on All Things Distributed, Amazon Chief Technology Officer Werner Vogels and Swaminathan 'Swami' Sivasubramanian, Vice President of Data and AI, unveiled AWS Emulator—a local, open-source mock runtime covering more than 200 AWS cloud services. What distinguishes the release is not merely its breadth across 400 public API models, but its genesis: the entire software package, along with its extensive test suites, was conceived, authored, and verified by autonomous AI coding agents.
The project originated from Matt Goldberg, a Senior Principal Engineer on the AWS Developer Tools team. Goldberg hypothesized that if an autonomous agent were supplied with machine-readable API specifications and access to production AWS infrastructure as an absolute oracle, it could iteratively build a conforming local emulation without manual human programming. Using Smithy—Amazon's open interface definition language describing operations, inputs, outputs, and type constraints—Goldberg directed the Kiro coding agent to construct an emulator for Amazon Simple Queue Service (SQS). In approximately 15 minutes, the agent returned a functioning, passing implementation.
To scale that prototype across hundreds of services, the AWS engineering team constructed a verification harness centered on fidelity tests. These end-to-end customer scenarios are compiled directly from Smithy definitions. The agents run each test against live AWS infrastructure to verify exact runtime behavior, including edge-case ordering and error serialization. Once validated against the cloud, the same test suite is executed against the local emulator; any deviation points directly to a missing feature or subtle behavioral mismatch that the agent must autonomously rectify.
Sivasubramanian drew a direct parallel to the early architecture of Amazon's foundational distributed services. Twenty years ago, building the replication engine for 'Alf'—the Paxos-based consensus locking service underlying Amazon S3 and DynamoDB—demanded four months of intensive manual systems engineering and custom JVM simulation. By contrast, when Sivasubramanian tasked Kiro with rebuilding that same replication layer given formal specifications and an automated test harness, the agent completed the task in two hours.
The emulator is published under the CC0 Creative Commons Zero dedication, placing the codebase squarely in the public domain. Amazon has made the entire commit history and agent-submitted pull requests accessible on GitHub. Crucially, the repository includes a dedicated 'things-that-do-not-work' directory, documenting exact operational discrepancies where current emulated behaviors diverge from live cloud infrastructure.
Every contribution to the codebase is screened by an automated merge gate consisting of deterministic static boundaries and specialized adversarial security agents. These AI reviewers systematically test prospective patches for injection vulnerabilities and unintended side effects before merging. Amazon confirmed that further architectural insights and developer tooling integrations will be showcased at the upcoming AWS re:Invent conference.