AWS Releases AWS Emulator Under CC0, Saying Agents Wrote It Entirely; Its Docs Say It Does Not Evaluate IAM Policies the Way AWS Does

Overview

AWS released AWS Emulator on November 11, 2026: a local program that serves an AWS API endpoint so the AWS CLI, the AWS SDKs and coding agents can be tested without calling AWS. The AWS News Blog, written by Daniel Abib, says it is generally available with no additional service fee, and Amazon’s blog post by Sr. Principal Engineer Matthew Goldberg says the emulator “was written entirely by autonomous agents.” The source is dedicated to the public domain under CC0, according to the project README. AWS’s documentation and several news outlets also describe limits, covered below. The Herald’s contributor model is built by Anthropic, which Goldberg’s post names in describing how the project began.

What was released

On Linux the emulator installs from an RPM or Debian package; on macOS and Windows it runs in Docker, and the AWS blog says AWS does not publish a ready-made image, so users build one from a Dockerfile in the documentation. The same post says it is written in Rust, serves HTTPS on 127.0.0.1:9297, requires every request to be signed with Signature Version 4 when it runs as a local process, and by default sends anonymous usage and error data that users can turn off. For tests in Rust, the README describes an embedded mode that runs one isolated instance per test.

The README also says the repository does not take pull requests and that building from source needs at least 100 GB of free disk for target/, with embedded mode’s first build on the same order. The New Stack quotes Tom “Spot” Callaway, a principal open source software strategist at AWS, on earlier emulators: “The challenge is that they get out of date very quickly.” The New Stack adds that the launch “may raise some questions about the future of LocalStack.” The Herald’s November 8 article on ended or frozen community editions covered LocalStack.

How AWS says it was built

Goldberg writes that researchers at Anthropic showing that agents could build a C compiler “was a lightbulb moment.” He then asked Kiro, which The New Stack calls AWS’s agentic coding tool, to build an Amazon SQS emulator, received a passable first version after 15 minutes, and had it build “the rest of AWS” over a weekend. Three teams of agents now write fidelity tests that are verified against real AWS, replay them to find divergences, and fix each filed issue behind a merge gate with security and regression reviews. Goldberg writes that his team has not written or reviewed any of the code. The AWS blog adds that an independent security agent reviews changes adversarially.

What AWS says it does not do

AWS’s documentation lists these limitations: the emulator does not evaluate IAM or resource-based policies the way AWS does, does not keep resources separate by Region or account, keeps all resources in memory and deletes them when it stops, and has not been tested for long-running use or load testing. Its fidelity page says “Today, the emulator matches AWS most closely for control plane APIs” and that data plane fidelity “and support for infrastructure as code tools such as AWS CloudFormation, Terraform, and the AWS CDK, varies by service.” It adds that an operation that is not implemented returns HTTP 501 with the error code NotImplemented.

What outlets and analysts say

InfoWorld quotes Ashish Chaturvedi, executive research leader at HFS Research: “Faster feedback in the inner development loop shortens delivery.” It quotes Devin Pratt, research director at IDC, saying that at launch the emulator “does not simulate throttling or outages, snapshot and restore, or per-Region availability, which matter most for resilience.” It also quotes Advait Patel, a senior site reliability engineer at Broadcom: “If your tests never see a 429 or a timeout, your retry and backoff logic never gets tested.”

SiliconANGLE quotes Holger Mueller of Constellation Research describing AWS’s answer as building “a digital twin of its public cloud.” The Register, in a piece by APAC Editor Simon Sharwood published at 05:47 UTC on November 12, writes that AWS’s intent “appears, as ever, to encourage more use of its cloud.”

Where the sources disagree

The Herald did not reconcile these differences.

• Unimplemented operations. The AWS blog and documentation say HTTP 501; InfoWorld says HTTP 503.
• Size of the supported surface. Goldberg’s post refers to 438 AWS API models; The New Stack says 438 models covering more than 200 services; InfoWorld writes that, according to AWS, the emulator “supports all 426 AWS services”; The Register says it found over 1,300 items on the API-models repository page. On November 12 the Herald counted 436 service-model JSON files in that repository through GitHub’s API.
• Number of agents. Goldberg’s post subheading says “Thousands of autonomous AI agents wrote every line of code in AWS Emulator,” and its body says hundreds work concurrently; SiliconANGLE says dozens.
• Why CC0. The New Stack writes that AWS was not sure it could copyright the agents’ work; Goldberg writes that the insights from the build history were the main motivation for the public dedication.

The Herald’s own tally of the scores file

The documentation says fidelity scores are published with each release. The v1.0.322 release, published 2026-11-11 00:32 UTC, lists fixes in its text and attaches scores.json.gz, which records a scoredAt time of 2026-11-11 00:09 UTC. Counting its units array, the Herald found 42,937 entries: 32,402 of kind “behavioral” (20,364 pass, 11,958 fail, 80 skipped), about 63% passing, and 10,535 of kind “wire” (3,879 pass, 6,656 fail). Among behavioral entries that name a single service, the Herald counted S3 at 763 of 924, DynamoDB 656 of 801, SQS 286 of 304, Lambda 293 of 616 and EC2 576 of 1,300 passing. These are the Herald’s tallies, not AWS’s published per-service scores. The documentation does not define “wire,” and the Herald did not map entries to the per-service scores AWS describes. The evaluation suite manifest listed 29,676 rows when read on November 12, against 32,402 behavioral entries in the scores file; the fidelity page says tests can be added, rebuilt or retired.

What is not known

None of the four outlets cited reports running the fidelity tests itself, and the Herald did not run the emulator. The fidelity page says each release names the fidelity-test version used; in the release text and the scores file’s top-level fields (commit, schemaVersion, scoredAt, units), the Herald found the emulator commit but no test version. AWS says it will share more about the project at its re:Invent conference, which The New Stack reports starts on November 30.