Amazon Open-Sources AWS Emulator, a Local AWS Stand-In Written Entirely by Agents
On this page 9 sections
TL;DR
AWS Emulator, released Nov 11, is a free program that emulates more than 200 AWS services on your own machine. Point the AWS CLI or any SDK at a local endpoint and your code runs against emulated resources: no AWS account, no network, no bill until you deploy for real. AWS says autonomous agents wrote every line of it, and that the engineering team "hasn't written or reviewed any of the code." Agents also maintain the thing, checking their work against the real cloud and merging thousands of fixes a day. The code is public domain under CC0, commit history included.
What shipped
The emulator is a Rust program that serves an AWS API endpoint on your machine, HTTPS on 127.0.0.1:9297 with a self-signed certificate. Set environment variables to point an SDK or the CLI at it, run your code as usual, and it talks to emulated resources instead of AWS. There is no account requirement and no service fee; you pay when you deploy on the real cloud. The user guide covers server mode, an embedded mode that compiles the emulator into Rust tests in-process, Docker images for Windows and macOS, and Linux packages.
Installation is agent-first:
aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1That needs AWS CLI 2.35.0 or later, and from there your coding agent sets the emulator up itself. The repository takes issues but no pull requests, because the fixers are internal agents running on Amazon Bedrock AgentCore with CodeCommit as their working repo; GitHub gets the code at the end of the release pipeline. Telemetry is on by default, and the README documents three ways to switch it off, starting with --telemetry off.
The swarm that maintains it
The origin was a weekend. Matt Goldberg, the senior principal engineer behind the project, had seen a demo where a team of Claude agents built a C compiler and took the lesson that agents given a well-defined spec and a reference implementation can copy the behavior faithfully. AWS owns both: the public AWS API models, Smithy specs describing every service's operations and errors, and the cloud itself. He gave Kiro a three-step prompt, got a passable Amazon SQS in 15 minutes, and had Kiro build the rest of AWS by Sunday. His team started testing their software against it that Monday.
The production version runs on three agent teams. Team one writes fidelity tests, end-to-end scenarios like creating a queue, sending a message, receiving it, deleting it, and each test must pass against real AWS before it counts. Team two replays those tests against the emulator and files failures, grouped by suspected root cause. Team three takes one issue per agent, fixes it, and pushes the change through a merge gate that checks compilation, coding standards, security, and fidelity regression.
"The emulator was written entirely by autonomous agents. Our team hasn't written or reviewed any of the code."
That is Goldberg, in the launch post. AWS says hundreds of agents work the codebase concurrently and thousands of issues get fixed every day. The repo passed 5,000 stars inside two days.
Coloring books for agents
Goldberg's launch post opens with him ruining a group drawing of dinosaurs in kindergarten, having arrived a year too early to hold a crayon properly. His classmates screamed at him. The problem, he decided, was competence, and his conclusion for agents is the same: "I think agents also need coloring books."
A coloring book hands you crisp outlines and a bounded region to fill in, and the lines stop you from wrecking the neighboring section. Here the outlines are the Smithy spec, the fidelity tests, and the merge gate, and inside them the agents fill in service behavior all day. Swami Sivasubramanian, AWS's VP of agentic AI, writes in a launch post on Werner Vogels' blog that when agents write the code and check their own work, "the engineer's job is to give them room to work on their own and to build the specs, tests, and merge gates that decide when that work is correct."
He learned that babysitting one. Twenty years ago he and Al Vermeulen spent about four months building the replication engine inside Alf, Amazon's Paxos-based lock service. This year he set Kiro on rebuilding it and supervised the agent for four nights before realizing it had no way to check its own results. Once it had a spec and a test environment, the rebuild took about two hours.
The honesty folder
Every issue the agents have filed is public, split into issues/things-that-now-work and issues/things-that-do-not-work. Fidelity scores ship with each release on the releases page, and the evaluation suite is a separate CC0 repo the agents also maintain: one Python file per test, read-only for humans, with every test verified against at least two sandboxed AWS accounts. Its README puts the limit in one line: "Behavior that no fidelity test covers is not verified."
The commit history holds the odd corners. In an earlier version, told to support AWS Lambda, the agents wrote their own Python interpreter rather than reuse one. Nobody asked for a programming language; the agents had a region to fill in, and they filled it. AWS has since told them to reuse existing libraries.
What it will not do
The emulator's own README rules out production workloads and security testing. IDC analyst Devin Pratt, in GeekWire's coverage, adds that it does not simulate throttling, outages, or region-by-region service differences, and it does not sandbox the code it runs. Building from source is Linux only and wants about 100 GB of disk for the Rust workspace, which has one crate per service. On Windows and macOS you run the Docker image instead.
The LocalStack angle
LocalStack has owned local AWS emulation for years. In March 2026 it consolidated its Community and Pro images into one distribution that requires an account and an auth token, and its free tier is now limited to non-commercial use. A CC0 emulator with no account requirement, shipped by AWS itself, changes the default for everyone who was pinning an old Community image to avoid the token. AWS says more is coming at re:Invent, and Goldberg promises a series of posts on the multi-agent setup.
Why builders should care
A coding agent works in a loop: write, run, read the result, decide, repeat. Against real AWS that loop burns latency and money on every iteration, and agents iterate constantly. The emulator puts the infrastructure inside the loop, which is why it exists at all; Goldberg's team was testing their own software against the weekend prototype by Monday morning.
The transferable part is the recipe, and you do not need AWS-scale resources to copy it: a written spec, a reference implementation to check against, tests verified on the reference before they count, and a merge gate with adversarial reviewers. Any API surface you own already has the first two.
If you run a homelab, this is an odd little gift: AWS-shaped services answering on your own iron, no egress, no account. Point Terraform at it and see what breaks.
Key Takeaways
- AWS Emulator (Nov 11) emulates more than 200 AWS services locally, built from 438 public Smithy API models, free under CC0 with the full commit history public. No AWS account, no service fee; telemetry is on by default with documented opt-out.
- Autonomous agents wrote every line, and the human team wrote and reviewed none of it. Three agent teams maintain it now: test writers verified against real AWS, issue finders that file failures by root cause, and fixers at one issue per agent behind a merge gate.
- Scale of the swarm: tens of thousands of fidelity tests, thousands of issues fixed per day by hundreds of concurrent agents, and 70 numbered releases between Nov 8 and Nov 12.
- It began as a weekend experiment: a three-step prompt got Kiro a passable Amazon SQS in 15 minutes and the rest of AWS by Sunday, and the team used it Monday.
- The limits are public: no throttling, outage or region simulation, no sandboxing, no production or security use. Fidelity scores ship per release, and the backlog sits in issues/things-that-do-not-work.
- The reusable pattern for any team: spec, reference implementation, verified tests, and a merge gate. Swami Sivasubramanian's Kiro rebuild of a four-month hand-built replication engine took about two hours once the agent had a spec and a way to check its work.
Sources: Matt Goldberg, How we built the agents that built AWS Emulator, About Amazon (Nov 11, 2026), Swami Sivasubramanian, All Things Distributed (Nov 11, 2026), aws/aws-emulator on GitHub, aws/aws-emulator-evaluation-suite on GitHub, AWS Emulator User Guide, GeekWire (Nov 11, 2026), LocalStack, The Road Ahead (Dec 18, 2025), Anthropic, agents build a C compiler