AWS Built a Free Stand-In for Its Own Cloud, and the Coders Were Robots
AWS Built a Free Stand-In for Its Own Cloud, and the Coders Were Robots
It started with a prompt and a Saturday. An AWS engineer named Matt Goldberg gave Kiro, Amazon's coding agent, a three-step instruction to build a working copy of Amazon SQS, the company's message queuing service. Fifteen minutes later, he had a passable first draft. By the end of the weekend, the agent had produced a rough version of much of the rest of AWS, with obvious holes. On Monday, his team started using it to test their software.
That experiment is now a product. AWS Emulator, released this week, simulates AWS services on a developer's own computer. It is built from more than 400 API models covering more than 200 services, and it lets people test code without a network connection and without running up a bill on the real cloud. The company is giving it away for free under CC0, a public domain dedication, with the full commit history published on GitHub.
The timing is deliberate. Coding agents such as Kiro and Anthropic's Claude Code can write software faster than teams can check it. Each time an agent writes code, it runs the code, reads the output, and tries again, sometimes dozens of times. When every one of those loops has to hit the real cloud, the delays and charges pile up. The emulator moves that loop onto the laptop, where it is fast and free.
Agents building the agents' playground
The irony is hard to miss. AWS used AI agents to build a tool for AI agents, and the agents keep it running too. The work is split across three teams, all running on Amazon Bedrock AgentCore. One team writes what the company calls "fidelity tests," realistic scenarios such as creating a queue, sending a message and deleting the queue, then checks each against the real AWS cloud. A second team runs those tests against the emulator and files issues for every failure. A third fixes them, with one agent assigned to each issue.
Before any change lands, it passes an automated gate, where other agents review coding standards and security. Goldberg says hundreds of agents now fix thousands of issues a day.
Goldberg described the agents in a blog post with an analogy that stuck. He compared them to himself as a kindergartner who ruined a group drawing of dinosaurs. The trouble wasn't bad intent, he wrote, just a lack of skill. The fix, for agents, is structure. "I think agents also need coloring books," he wrote.
That idea runs through AWS's broader pitch. Swami Sivasubramanian, the company's vice president of agentic AI, said that when he tried to have Kiro rebuild a replication engine he and a colleague had written by hand nearly twenty years ago, the agent spent four nights struggling. Once he gave it a spec and a way to test its own work, the job took about two hours.
The engineer's job shifts from typing code to writing the specs, tests and merge rules that decide whether the machine's output is correct.
The uncomfortable part for LocalStack
Any AWS emulator lands in a crowded market, and the most prominent competitor is LocalStack. It built a business on the same promise, offering developers a local version of AWS. This year, it moved its free version behind an account requirement and limited that free tier to noncommercial use, which means companies with commercial work have to pay.
AWS's move changes that math. A free, public domain emulator from the company that owns the cloud is a hard offer to ignore, especially for individual developers and small teams who were already weighing the cost of LocalStack's restrictions. It doesn't mean LocalStack loses overnight, since established tools carry habits, integrations and institutional trust. But it puts pressure on a company whose main selling point just became available for nothing.
There is also a question about what "free" means in practice. AWS's emulator is open under CC0, which lets anyone use, change or redistribute it without asking permission. LocalStack's model depends on accounts and licensing. Developers will weigh the convenience of a no-strings tool against the support and polish a commercial vendor provides.
The gaps are the story
The emulator is not a finished replica. An IDC analyst, Devin Pratt, told InfoWorld that it does not yet simulate throttling, outages or the regional differences in which services are available. It also does not sandbox the code it runs, which matters when the code comes from an agent. AWS has been unusually open about the flaws, publishing the list of known gaps in a GitHub folder called "things-that-do-not-work."
That folder may be the most revealing part of the project. It is a public admission that the software was built by machines that learn by trial and error, and that the humans overseeing it are relying on tests to tell them when the work is right. AWS says it will share more at its upcoming re:Invent conference. Whether the emulator becomes a dependable part of everyday development or a clever experiment that stays in the tutorials will depend on how many of those gaps the agents can close, and how fast.
Aaron Rose is the founder and publisher of Tech Reader Magazine, covering system architecture, cloud platforms, and AI policy.