Subscribe
0%
Illustration for: AWS Emulator gives you 200+ AWS services on your laptop for free, and agents wrote all of it. What it does and where it falls short

AWS Emulator gives you 200+ AWS services on your laptop for free, and agents wrote all of it. What it does and where it falls short

AWS released a free, offline, CC0-licensed emulator built from 400+ public API models, with code AWS says was written and tested entirely by AI agents. It's useful for local tests and coding agents, but it doesn't enforce IAM policies, keeps everything in memory, and fidelity varies by service.

Listen to this article0:00 / 4:28

Amazon Web Services released AWS Emulator on November 11: a free local stand-in for more than 200 AWS services that developers and coding agents can call without an AWS account or network connection. The unusual part is how it was built. In a guest post on Werner Vogels' All Things Distributed, AWS's Swami Sivasubramanian says the code was written and tested entirely by agents.

That matters for two groups. Anyone who tests AWS code against real accounts, or pays for third-party emulators, now has a first-party option at no cost. And for anyone tracking how far AI coding has come, this is one of the largest agent-written codebases a major vendor has shipped as a product, with its test corpus and issue backlog in public.

What you get

Item Detail
Coverage 200+ services, generated from 400+ public Smithy API models (Unite.AI puts it at 438)
Price Free. You still pay for anything you deploy to real AWS
License CC0 public-domain dedication; vendored framework code is Apache 2.0
Platforms RPM and Debian packages on Linux; a container on macOS and Windows; an embedded mode for Rust tests
Endpoint HTTPS on 127.0.0.1:9297 with a self-signed certificate, per MiniStack's walkthrough of the README
Telemetry On by default; turn off with --telemetry off or EMULATOR_TELEMETRY=off
Contributions Bug and fidelity reports through GitHub Issues; pull requests aren't accepted

According to Unite.AI, the AWS CLI, SDKs and coding agents connect through environment variables, so application code doesn't change. Operations that aren't implemented return HTTP 501 rather than failing silently. Developers Digest reports that coding agents can add it as a skill with the AWS CLI (version 2.35.0 or later):

aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1

How agents built it

Sivasubramanian says the idea came from Matt Goldberg, a senior principal engineer on AWS's developer tools team. The spec is AWS's public API models; the reference implementation is real AWS. Agents first tune each end-to-end "fidelity test" until it passes against the real cloud, then build the emulator until it passes the same test. The first prototype, an SQS emulator, came out of the Kiro agent from one prompt in about 15 minutes, he writes.

Unite.AI describes three agent teams: test writers, issue finders that group related failures, and fixers that each take one issue. Goldberg's team reports tens of thousands of fidelity tests and thousands of issues fixed a day by hundreds of concurrent agents. Those are AWS's own figures. Every change has to pass a merge gate covering the build, coding standards, adversarial security review by separate agents, and a fidelity regression check. Goldberg says his team hasn't written or reviewed any of the code; humans built the specs, test machinery, gates and release pipeline.

Developers Digest counted 29,684 commits in the repo, the first dated October 21.

Where it falls short

AWS is explicit about the limits, per the README summaries in Unite.AI and MiniStack:

  • No real IAM. It doesn't evaluate IAM or resource-based policies the way AWS does, so a request AWS would deny can succeed locally. Keep permission and security tests on real AWS.
  • One account, one Region. It accepts any Region or account ID but doesn't separate resources by them.
  • In memory only. Resources vanish when the emulator stops. It's for test runs, not long-running or load testing.
  • Uneven fidelity. Each release publishes per-service pass rates, and untested behaviour isn't verified. A folder called things-that-do-not-work lists known mismatches.
  • Heavy Rust builds. The README recommends at least 100 GB of free disk when building the embedded mode from source, according to Developers Digest.

On Hacker News, where the launch drew roughly 795 points, commenters questioned whether a commit count near a thousand a day says anything about maturity, and whether AWS will keep maintaining it. An AWS employee replied that it is a real product. One commenter pointed to the open-source Floci as an alternative that worked well for them.

Who should try it

Teams that already mock AWS in CI, and anyone letting coding agents write infrastructure code, are the obvious fit: an agent can create a resource, inspect it, and fix its code without waiting on a deploy. Before switching, check your services' pass rates in the latest release, keep IAM tests on a real account, and turn telemetry off if your policies require it. AWS says more details are coming at re:Invent.

Sources

  1. All Things Distributed — AWS Emulator and the engineer's job when agents write the code allthingsdistributed.com
  2. Unite.AI — AWS Releases Local Emulator Written Entirely by Autonomous AI Agents unite.ai
  3. Developers Digest — AWS Emulator: Free Local AWS Built by AI Agents (HN Review) developersdigest.tech
  4. MiniStack — AWS launches AWS Emulator ministack.org

#AWS#AWS Emulator#Kiro#coding agents#local development#testing#open source