1. Home
2. News
3. AI products

AWS releases a free emulator of 200 services, and says AI agents wrote every line

One engineer's weekend prompt to Kiro became a product that hundreds of agents now build, test and maintain on their own.

By Think Facility · November 11, 2026 · Amazon AI agents Save

On November 11, 2026, Amazon Web Services released AWS Emulator, a free program that copies more than 200 AWS services onto a developer's own machine so code can be tested without touching the cloud. The launch post says autonomous AI agents wrote all of it.

The team behind it says it didn't check their work line by line either. "Our team hasn't written or reviewed any of the code," the post says. Agents keep it up to date now, testing it against real AWS and fixing what doesn't match.

What

AWS Emulator, a local emulator of AWS services

Covers

more than 400 API models covering more than 200 AWS services

Written by

autonomous agents, with no human review

License

CC0, a Creative Commons dedication to the public domain

Runs

free and offline on your own machine

How the agents built it

It started as a weekend experiment. Matt Goldberg, a senior principal engineer on the AWS Developer Tools team, gave Amazon's coding agent Kiro a three-line prompt: write tests from the AWS command line tool, fix them until they pass against real AWS, then build an in-memory copy that passes the same tests. About 15 minutes later Kiro handed back a passable first version of Amazon SQS, AWS's message queue service.

Goldberg says he refined the prompt and had Kiro build "the rest of AWS over the weekend." It had gaps, but his team was testing their own software on it by Monday. Turning that into something customers could use took much longer, and his post spends most of its length on the guardrails.

Three teams of agents do the work. The first writes fidelity tests from AWS's public API models and checks each one against the real cloud before it's accepted. There are tens of thousands of them. The second team replays those tests against the emulator and files an issue for each failure. The third takes one issue each, fixes it and submits the change.

Every change then passes through a merge gate. That gate is more agents: some check coding standards, some check security, and some make sure the change didn't break anything that used to work. Goldberg writes that thousands of issues get fixed a day with hundreds of agents working at once, and that "Without this merge gate, we wouldn't trust what our agents are building."

What changes for the engineer

Amazon put a second post up the same day on the blog of its CTO, written by Swaminathan Sivasubramanian, who started at Amazon as an intern on that CTO's team in 2004. He frames the emulator as an answer to one problem: agents now write code far faster than any team can check it against production.

His own test of the idea is the most concrete number in either post. About twenty years ago he helped build the replication engine inside Alf, Amazon's internal lock service, and it took about four months. He tried rebuilding it with Kiro and supervised the agent for four nights before he saw it had no way to check its results. With a spec and a test environment, the rebuild took about two hours.

I'd read that as a plan for how Amazon wants to work, and it comes with its own caveat. Sivasubramanian also writes that they're early in learning how engineers drive teams of agents, and that they expect to keep adding guardrails.

What it means for people who test software for a living

The job AWS handed to agents here is testing, and testing is a large US occupation. The Bureau of Labor Statistics counts about 187,600 jobs for software quality assurance analysts and testers in 2025, at a median annual wage of $104,300 in May 2025. Their listed duties include creating test plans and reporting defects, which is close to what the first two agent teams do.

The bureau's projections don't show those jobs shrinking yet. It expects testers to grow 6% from 2025 to 2035, and software developers 10%, against 3% for all occupations. Those projections were made before tools like this shipped, and one emulator built inside AWS doesn't tell anyone how many companies will copy the approach. What it does show is a large company deciding that agents checked by other agents are good enough to put in customers' hands.

The code is public. Every issue the agents filed is in the GitHub repository, split into a folder of things that now work and a folder called things-that-do-not-work. AWS says it'll share more about the project at re:Invent.