AWS offers developers a free local replacement for its cloud, built from scratch by AI agents – GeekWire
AWS claims its engineers haven’t written or reviewed a single line of code in its new developer tool, and the company presents this as a selling point.
AWS Emulator, released Wednesday, simulates AWS services on a developer’s own machine. It is built from over 400 API models covering over 200 AWS services. Developers and AI coding agents can use it to test code without a network connection and without paying for the real cloud.
The tool solves a problem that coding agents like Amazon’s Kiro and Anthropic’s Claude Code have made worse. Agents can write code faster than teams can test it on the real cloud. Agents verify their work by running the code and reading the results, over and over again, and doing this in the cloud results in delays and costs.
The emulator allows them to run this loop on the developer’s own machine.
Ironically, or perhaps fittingly, AWS used AI agents to create the emulator. Agents also maintain it, test it against real AWS, and fix any gaps they find.
“The emulator was written entirely by standalone agents. Our team did not write or review any code,” Matt Goldberg, the AWS senior principal engineer behind the project, wrote in a blog post.
Goldberg likened his agents to himself as a kindergartner, too young to hold a pencil properly, who was yelled at by his classmates for ruining a drawing of a group of dinosaurs. The problem was not bad intent, he wrote, but simply a lack of skill. The answer, for agents, lies in clearly defined areas to fill.
“I think agents need coloring books too,” he wrote.
Agents work off AWS’s public API models, which describe the behavior of each service and verify their work against the real AWS cloud.
Agents run on Amazon Bedrock AgentCore in three shifts. One of them writes “fidelity tests” – real-world scenarios such as creating a queue, sending a message, and deleting the queue – and verifies each of them against a real AWS. A second runs these tests on the emulator and logs issues if they fail, and a third fixes them, one agent per issue.
Each change goes through an automated portal where other agents check coding standards and security. Goldberg wrote that hundreds of agents now resolve thousands of issues a day.
The project began as a weekend experiment, inspired by Anthropic’s demo earlier this year in which a team of Claude agents built a C compiler. Goldberg gave Kiro a three-step prompt and had a passable first version of Amazon SQS, AWS’s message queuing service, in 15 minutes.
By the end of the weekend, Goldberg wrote, Kiro had built a draft version of the rest of AWS, with obvious gaps, and his team began using it to test his software that Monday.
Swami Sivasubramanian, AWS vice president for agentic AI and emerging technologies, linked the project to a lesson he described to GeekWire in June. He spent four nights babysitting Kiro as he attempted to rebuild a replication engine that he and Allan Vermeulen had built by hand almost 20 years ago.
Once he gave the agent the specifications and a way to test his work, the job took about two hours.
“When agents write the code and check their own work, the engineer’s job is to give them the opportunity to work alone and create the specifications, tests, and merge gates that decide when that work is correct,” Sivasubramanian wrote in a guest post on CTO Werner Vogels’ blog.
It’s the same idea behind Amazon’s recent decision to rewrite its engineering playbook for an era where AI writes the code.
The emulator has limitations. Devin Pratt, an IDC analyst, told InfoWorld that it doesn’t yet simulate limitations, outages, or differences in available services and features by region, and it doesn’t sandbox the code it runs. Agents’ backlog of known deficiencies is public, in a GitHub folder called “things that don’t work.”
There are other AWS emulators, including LocalStack, which this year put its free version behind an account requirement and limited it to non-commercial use. AWS has released its emulator for free under CC0, a dedication to the public domain, with the full commit history on GitHub.
The company says it will share more at its upcoming re:Invent conference.
Gn bussni
