AWS Emulator: thousands of AI agents wrote every line, Amazon says

The latest:

Amazon has published AWS Emulator, a downloadable binary that reproduces AWS services on a developer’s own machine, and says thousands of autonomous agents wrote every line of its code. Writing on About Amazon on 11 November 2026, Sr. Principal Engineer Matt Goldberg said the team neither wrote nor reviewed any of it. The code sits on GitHub under a CC0 license.

Details:

• The product: AWS Emulator runs locally so developers can build and test AWS applications without calling the real cloud. In a guest post on Werner Vogels’ All Things Distributed blog on 11 November, Swami Sivasubramanian said it covers more than 200 AWS services, runs free and offline, and is built from more than 400 API models written in Smithy.
• The claim: Goldberg wrote that thousands of autonomous agents produced the entire codebase, stating: “Our team hasn’t written or reviewed any of the code.” The repository, aws/aws-emulator, is updated continuously, and a separate repository, aws/aws-emulator-evaluation-suite, holds the conformance test suite.
• The method: Goldberg said the approach needs a precise specification and a reference implementation — here the AWS API models and the real AWS cloud. He gave the Kiro tool simple instructions: write tests based on the AWS CLI, iterate until they pass against real AWS, then build an in-memory implementation that passes them too.
• The first run: Kiro produced an initial version of Amazon SQS in roughly 15 minutes and the rest of AWS over a weekend, Goldberg wrote, with obvious gaps. An early scaffold handled 438 API models without implementing anything, and turning that draft into a durable product took considerably longer.
• Three agent teams: Test writers build conformance tests from public API models and validate them against real AWS, numbering in the tens of thousands. Failure finders rerun those tests against the emulator, log failures and cluster ones likely sharing a single cause. Fixers each take one defect, trace its cause and amend the code.
• The merge gate: Changes pass a gate checking build integrity, coding standards, security and conformance regressions, with anything failing sent back for rework. Sivasubramanian said the gate includes several security agents running adversarial reviews alongside a general code reviewer.
• The scale: Goldberg said thousands of defects are fixed daily and hundreds of agents run at once. Changes merge into the main branch on AWS CodeCommit, the agents’ primary repository, then are compiled, built and tested before a release branch, packaged as a binary and published to GitHub at the final stage.
• The speed claim: Sivasubramanian said rebuilding a replication engine took about four months the first time and roughly two hours with agents, given a specification and a test environment. An internal version of the emulator was completed over a weekend and spread across teams within a month.
• What is public: Sivasubramanian said the public repository carries the full change history and the defects agents logged. It holds two folders, issues/things-that-now-work and issues/things-that-do-not-work, the latter covering what the emulator still fails to do.
• Guardrails: In an earlier iteration the agents built their own Python interpreter to support AWS Lambda, prompting the team to instruct them to reuse existing libraries instead. Goldberg likened what agents need to a coloring book: defined spaces and boundaries. The idea was inspired by Anthropic research showing agents could build a C compiler themselves.

Between the lines:

The model Goldberg describes shifts human effort from writing code to defining the specification and the reference the agents measure themselves against — the API models and real AWS supply both. The coloring-book analogy and the Lambda interpreter episode point the same way: the agents’ output was only usable once the boundaries were drawn tighter.

What’s next

Goldberg promised a series of posts on multi-agent architecture, conformance measurement, quality standards and security. Sivasubramanian said AWS will share more at re:Invent. The public repository’s things-that-do-not-work folder tracks remaining gaps.