AI
AWS Releases Emulator Built Entirely by AI Agents for Local API Testing
AWS has open-sourced an emulator written in Rust by coding agents that lets developers test against 200+ AWS services locally without provisioning cloud infrastructure.
InfoQ
AWS has unveiled AWS Emulator, an open-source local testing tool written entirely by AI agents that provides a functioning HTTPS endpoint for over 200 AWS service APIs. The tool eliminates the need to spin up cloud resources or pay for service charges while iterating on code.
How It Works#
Developers can redirect their AWS CLI and SDKs to the emulator using simple environment variables like AWS_ENDPOINT_URL, with no changes required to application code. The implementation comes in multiple forms: Linux users can install it via RPM or Debian packages for standard CPU architectures; macOS and Windows users can run it via a lightweight Docker container. For Rust-based testing, an embedded mode allows the emulator to run directly inside the test process.
Built in Rust and generated from over 400 Smithy API models that AWS publishes for its SDKs, the emulator also includes a skill for coding agents that can start the tool and validate projects against fidelity benchmarks.
Agent-Driven Development#
The most distinctive aspect of AWS Emulator is that AI agents handled its entire development lifecycle. According to Matthew Goldberg, a senior principal engineer at AWS, three autonomous teams managed the work: one team generates fidelity tests from AWS API specifications and validates them against actual AWS; a second replays those tests and reports bugs; and a third implements fixes. All changes must pass a series of checks—compilation, code standards, security scanning, and fidelity regression testing—before merging. The infrastructure runs on Amazon Bedrock AgentCore Runtime Instances alongside AWS CodePipeline and CodeCommit.
Fidelity and Limitations#
AWS publishes versioned fidelity scores for each release, with tests passing only when API calls exactly match AWS behavior, including response codes, side effects, and error messages. Unimplemented operations return HTTP 501 with a NotImplemented error code.
Fidelity performance varies considerably by service. Control-plane APIs tend to be most accurate, while data-plane behavior and support for infrastructure-as-code tools like CloudFormation, Terraform, and AWS CDK differ widely. Key limitations include:
- No IAM or resource-based policy evaluation (unsuitable for security testing)
- Single shared account and Region per process

