Build a local AWS API emulator by letting an agent write tests against real AWS

AWS engineer Matthew Goldberg had the Kiro autonomous agent write CLI tests against real AWS, then code that passed them, producing a local AWS emulator with no human-written code.

Done withKiroinAWSDocker· Nov 16, 2026

What they did

Matthew Goldberg, a Sr. Principal Engineer at AWS, started the project as a weekend experiment, prompted by Anthropic's demonstration of agents building a C compiler on their own. He gave the Kiro autonomous agent a clear specification and a short three-step instruction: write tests that exercise the AWS command line interface, iterate on those tests until they pass against the real AWS cloud, then build a local implementation that passes the same tests. The result is a local service that answers many AWS API endpoints so developers can test code on their own machine, a dev server or in CI without calling AWS. It ships through GitHub, runs in Docker on Mac and Windows, and as RPM or Debian packages on Linux.

How it went

According to TechRadar's report, a first passable version came out in about 15 minutes. After refining the prompt over the weekend, Goldberg had a version his team could use by Monday. AWS says all of the code was written by agents and none of it was reviewed by humans on the team. Goldberg wrote that turning it into something fit for customers took far more than a weekend, mostly in building guardrails that keep the agents inside the team's rules. Not every AWS feature is covered.

Worth knowing

The key move is making the agent prove its tests against the real service first, so the tests become the spec for the copy. That costs real AWS calls during the test phase. This is a press write-up of an AWS team's own project, and the code is unreviewed by humans, so treat the emulator as a test aid, not a faithful copy of AWS.

Try it

With Kiro· 7 parts to fill in

I want a local fake of [service or API] so I can test without calling the real thing. Work in three phases. 1) Write tests that call [the service's CLI or API] for [list of operations], and run them against the real service in [my test account] until they all pass; do not change a test just to make it pass, fix it to match what the real service does. 2) Freeze those tests as the spec. 3) Build a local server in [language] that serves the same endpoints, and iterate until every frozen test passes against it. Ask me before any call that could cost more than [amount] or touch data outside [test account]. Stop when the local server passes the full suite and give me a list of what is not covered.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related