Amazon says agents wrote AWS Emulator and the team did not review the code, and Swami says an engineer built the scaffolding
Matthew Goldberg’s post, dated 11 November, puts the emulator under a CC0 licence and says three teams of agents test, file and fix it. Swaminathan Sivasubramanian’s guest post the same day says the release covers more than 200 services, and that a person built the runner the agents work inside.
Feature4 minIn-depth
James Okonkwo4 min read
- Amazon
- AWS
- Agents
- Kiro

Photo: AmazonSource page
Amazon on Wednesday released AWS Emulator, a local stand-in for AWS that a developer can run while testing an application, and two of its own posts do not describe the human role the same way. Matthew Goldberg, a senior principal engineer, writes on About Amazon that the emulator was written entirely by autonomous agents and that the team has not written or reviewed any of the code. Swaminathan Sivasubramanian, in a guest post on Werner Vogels’s blog the same day, writes that all of the emulator’s code was written and tested by agents, and then that an engineer built the scaffolding those agents run in. Vogels’s introduction says the emulator launched that day. The code is under a CC0 dedication in both accounts, and both point at a public GitHub repository.
Goldberg’s origin story is a prompt to Kiro. He told it to write tests from the AWS command-line interface, iterate until those tests passed against real AWS, and then build an in-memory copy that passed the same tests. He writes that Kiro returned a passable first version of Simple Queue Service in 15 minutes, and that a refined prompt produced the rest of AWS over a weekend, with gaps. Sivasubramanian tells the same 15 minutes and the same first service, and he adds the format of the spec: public API models written in Smithy. He also writes that an internal version came together over a weekend and started to be adopted within a month, and that turning it into something he would call robust took longer than that. Goldberg’s line on the same point is that a weekend result was not yet something he would hand a customer.
Three teams, two counts of the surface area, and a folder of failures
The working arrangement in Goldberg’s post is three groups of agents. One writes fidelity tests from the public API models and checks each test against real AWS before it is kept. He says there are tens of thousands of those tests. A second group replays the tests against the emulator and files issues. A third group takes one issue, edits the code, and has to pass a merge gate: the change must compile, pass an agent review against coding standards, pass a security check, and not make an existing test worse. If it fails, it goes back. He writes that hundreds of agents have been fixing thousands of issues a day. The repository, he says, keeps fixed issues under “issues/things-that-now-work” and the backlog under “issues/things-that-do-not-work.” Sivasubramanian describes the same gate and names the second folder. He does not print the daily issue count or the headcount of agents. He does say the gate includes several security agents looking for different ways to attack a change.
The size of the emulator is where a reader has to keep the sentences apart. Goldberg writes that the system began as an empty web server that could accept requests for 438 AWS API models and did nothing with them. That 438 is the count on the empty server, not a sentence about the build Amazon is shipping. Sivasubramanian writes that the release is built from more than 400 API models covering more than 200 services, that it runs free and offline on the user’s machine, and that AWS today has over 200 services, which is why a hand-built fake of each one was not going to get written. Four hundred and thirty-eight is more than 400. The two posts are not offering two inventories so much as a starting scaffold and a description of the result. Only Sivasubramanian states the service count.
Our team hasn’t written or reviewed any of the code.
— Matthew Goldberg, Amazon, in his post on AWS Emulator
Goldberg’s own history of the project is the place the “no review” line has a footnote he does write. An earlier run of the agents, he says, decided to build a Python interpreter so Lambda functions would run, and the team later told them to reuse libraries instead. That guidance is a human instruction about what the agents are allowed to invent. Sivasubramanian’s version of the same idea is broader. He writes that when agents write the code and check it, the engineer’s job is the spec, the tests and the gate, and he describes a personal trial: rebuilding a replication engine that had taken him and a colleague about four months, which an agent finished in about two hours once it had a spec and a test environment. He also says Amazon will talk about the work further at re:Invent. He does not give the day.
Simon Willison’s note, timestamped 9:18 p.m. on 11 November, is the account that puts a number on the commit log and a limit on the install. He writes that the repository shows 29,684 commits in the first three weeks of the project, and that macOS and Windows currently require Docker. He wants a single binary. Neither Amazon post prints the commit count or the Docker requirement. Goldberg says the release pipeline packages a binary for download and publishes the source to GitHub only at the end, after commits have been stacked, rebuilt and retested. What a developer can hold today, on the two Amazon accounts, is a local emulator whose failures are listed in public, whose tests are checked against the real cloud, and whose authors are agents operating inside a gate a person designed.


