426 services, 0 humans: AWS's emulator is written by agents
AWS launched a local emulator of its own cloud today, covering 426 services — and says autonomous agents wrote every line of the code, with no human reviewing it. I downloaded the release's own fidelity file: 56.5% of its 42,937 test units pass.
What launched
AWS Emulator is a local program that serves the AWS APIs on your machine, announced today (November 11) in a guest post on All Things Distributed by Swami Sivasubramanian. The source is public on GitHub at aws/aws-emulator, released under CC0 — effectively public domain. It is free, runs offline, and needs no AWS account. Point the AWS CLI or an SDK at it and your code runs against emulated resources instead of real ones.
The numbers AWS puts on it: more than 400 API models covering more than 200 services, per the launch summary — InfoWorld reports 426 services supported. It ships as an RPM or Debian package on Linux (the x86_64 RPM is 159 MB) and as a Docker image you build yourself on macOS and Windows. It serves HTTPS on 127.0.0.1:9297 with a self-signed certificate.
The part that actually matters: nobody human wrote it
Every line of the emulator was written and tested by agents, according to Amazon. The loop works like this: agents derive fidelity tests and end-to-end scenarios from the public Smithy API models, correct them until they pass against real AWS, then build the emulator until it passes the same tests. Every change goes through a merge gate that includes security agents doing adversarial reviews. Matthew Goldberg, a senior principal engineer on the project, says the team has not written or reviewed any of the code.
The commit history backs up the "agents did this" framing in a way a press release can't. The repo was created September 11 — 61 days before launch — and by this morning it had 29,684 commits and 322 releases. That is roughly 490 commits and 5 releases a day, every day, for two months. No human team ships at that cadence on a 400-service surface. The release notes read like machine output too: the launch release, v1.0.322, is mostly a long list of fidelity fixes with names like dynamodb-validationexception-messages-diverge-from-the-wording-aws and s3-listobjectversions-omits-markers-owner-and-checksum-members.
I checked the fidelity file it ships
AWS publishes versioned fidelity scores with each release, and the launch release ships a scores.json.gz. I downloaded it and counted, rather than taking the launch copy at face value:
| Measure | Count | Share |
|---|---|---|
| Test units, total | 42,937 | — |
| Units passing | 24,243 | 56.5% |
| Units failing | 18,614 | 43.4% |
| Units skipped | 80 | 0.2% |
| Individual steps passing | 49,810 / 68,007 | 73.2% |
| Services scored in the file | 355 | — |
Two things fall out. First, the honest launch number is 56.5%, not "AWS, on your laptop." A bit over half the scenario tests pass; step-level fidelity is better at 73.2%, which is the more useful figure — most scenarios get most of the way there and fail on one step. Second, the scores file covers 355 services against the 426 AWS claims, so about 70 services have no published score yet. That gap is worth watching, not assuming away.
One design decision deserves credit: operations that are not implemented or verified return HTTP 503 instead of fake success. An emulator that silently returns plausible-but-wrong data is worse than no emulator; a 503 is a loud, greppable failure. AWS also publishes the gaps in the open — issues/things-that-do-not-work/ lists behaviors that don't match AWS yet.
The limits AWS lists itself
This is a test tool, not a cloud: one region and one account (it accepts any region and account ID but doesn't keep resources separate), no IAM policy evaluation the way AWS does it, all state in memory and deleted when it stops, behavior checked against us-west-2 only (global services in us-east-1), no throttling or outage simulation, no snapshot and restore, and not tested for long-running use or load testing. And one flag for the privacy-minded: telemetry is on by default. --telemetry off or EMULATOR_TELEMETRY=off turns it off, but "runs offline on your machine" and "phones home by default" sit awkwardly in the same README.
Why now, and who it threatens
The timing is not subtle. LocalStack's Community Edition — the default answer to "how do I test AWS code without touching AWS?" for years — was archived in March 2026, and the unified image now refuses to start without an auth token tied to a registered account. Floci filled that hole: an MIT-licensed emulator, 128 services, past 31,000 GitHub stars by early November. AWS launching its own free, no-account emulator with a claimed 426 services is the incumbent answering both of them at once — with three times Floci's service count and the one thing neither rival has, which is AWS's own fidelity measurements against real AWS.
I didn't boot the emulator for this piece — the RPM is 159 MB and the from-source build wants 100 GB of disk for its per-service crates, so the scores file was the check I could actually run and stand behind. The question the launch really asks is not whether the emulator works today at 56.5%. It is whether you trust test infrastructure whose code no human has read, where the quality argument is "security agents review it" and the evidence is a public score file. AWS's answer is to publish the scores and the failing-issue list in the open, which is the right move — and also the reason the 56.5% is on the record at all. I would run it in CI with the 503s treated as test failures. I would not yet treat a green run against it as proof my code works on AWS.
Sources: AWS Emulator repository and v1.0.322 release, github.com/aws/aws-emulator (release published 2026-11-11T00:32Z; scores file scored 2026-11-11T00:09:58Z, downloaded and counted by me — 42,937 units, 24,243 pass, 18,614 fail, 80 skipped, 49,810/68,007 steps pass, 355 services); ministack.org, "AWS launches AWS Emulator" (November 11, 2026); InfoWorld, "AWS unveils local emulator service to help enterprises accelerate app development and testing" (November 11, 2026); The Circuitry, "AWS releases a local emulator written entirely by AI agents" (November 11, 2026); Floci/LocalStack background from AI Frontier Post (November 2026). Commit, release, star, and package-size figures are my own checks against the GitHub API and the AWS downloads server on November 11, 2026.

