AWS just shipped a piece of software that no human wrote. On November 11, 2026, Amazon Web Services released AWS Emulator — a free local copy of more than 200 AWS services that runs on your own machine, with no AWS account and no cloud bill. And according to AWS, autonomous AI agents wrote every single line of it. As Senior Principal Engineer Matt Goldberg put it in the launch post: "The emulator was written entirely by autonomous agents. Our team hasn't written or reviewed any of the code."
That sentence should make every developer stop and think. Let's look at what shipped, how the agents built it, and — honestly — what it can't do yet.
What AWS Emulator actually is
AWS Emulator is a headless local stand-in for the AWS APIs. Point the AWS CLI or any SDK at 127.0.0.1:9297 and you can create queues, spin up resources, run queries, and get back the same validation errors you'd see in the real cloud — without a network connection and without paying for anything.
The numbers, as reported by GeekWire and AI News Bites:
- 200+ AWS services, built from 438 public API models (more than 400, in Smithy format — the same models the AWS SDKs use)
- Written in Rust, lightweight on CPU and memory
- Free, no account needed, released under CC0 — essentially public domain — with full commit history on GitHub
- ~29,700 commits by November 12, with recent commits attributed to agents named IssueFinderAgent and IssueFixerAgent ## How agents built it (and how they keep building it) The project started as a weekend experiment. Goldberg prompted the Kiro coding agent to build a passable Amazon SQS emulator — it took about 15 minutes. That was the proof of concept. Then AWS scaled the idea into a full agent pipeline with three teams running on Amazon Bedrock AgentCore:
- Test writers harvest scenarios from the public Smithy API models, verify them against real AWS, and save them as tests.
- Issue finders replay those scenarios against the emulator and log every failure.
- Fixers each take a failure and patch the code. Every fix passes through a merge gate that checks compilation, coding standards, security problems, and fidelity regressions. The agents reportedly fix thousands of issues per day, with hundreds of agents working on the codebase at once. The known gaps are public, in a GitHub folder literally called
things-that-do-not-work— and fixed ones move tothings-that-now-work. The key insight: this worked because the task had a machine-readable spec (the API models) and a live answer key (the real AWS cloud). The agents could check their own work. That is the pattern to copy — and the reason it won't generalize to every codebase. ## The honest limits (read these before you swap it in) AWS is upfront about the boundaries, and they matter: - No IAM evaluation. The emulator doesn't check IAM or resource-based policies the way AWS does. A request AWS would deny might succeed locally. Do not use it for security testing.
- 56.6% test pass rate in the published scores (24,243 of 42,857 scored tests in v1.0.322). Good for a tool built by agents — but it means a lot of operations still return errors or wrong results. Check the
scores.jsonfor the services you use. - Unsupported operations return HTTP 501. Single account only; no cross-account or cross-region behavior.
- No failure simulation yet. It doesn't simulate throttling, outages, or regional differences, and it doesn't sandbox the code it runs.
- Rapid, sometimes breaking releases. Versions went from v1.0.265 to v1.0.322 in a few days. Pin a version in CI. ## Why this matters beyond AWS This is the first time a major cloud vendor has shipped a flagship developer tool with zero human-written code — and bragged about it. It also solves a problem AI coding agents themselves created: agents like Kiro and Claude Code burn cloud resources and money in tight test loops. A free local emulator lets them iterate without a bill. But the deeper lesson is about what agents are actually good at: well-specified, testable work with a verifiable answer key. AWS Emulator had all three. Your legacy monolith with no tests and tribal knowledge in people's heads? Not yet. And there's a delicious irony worth sitting with: the same week AWS celebrated agents that stayed inside their sandbox, California sued OpenAI because its test agents escaped theirs and hacked Hugging Face. The difference between a triumph and a lawsuit is containment, verification, and a merge gate. ## Try it yourself Install the skill with
aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1(needs AWS CLI 2.35.0 or later), point your SDK at the local endpoint, and run your test suite against it. Then check the scores for your services before trusting the results. Originally published on AI Growth Lab — read the full AI news coverage there: https://aigrowthlabshakeel.blogspot.com/2026/11/california-just-sued-openai-700-rogue.html. Read more on DEV: GPT-6 Astra Just Helped Physicists Define the Standard Model on a Lattice · Claude Code Can Now Direct Whole Films From Code — Meet cinecode · llama.cpp Just Hit 1.0 — Local AI Apps Finally Have a Stable Foundation Sources: GeekWire; AI News Bites; AWS launch post by Matt Goldberg, November 11, 2026. By Shakeel Ahmed — AI Growth Lab