AWS releases local cloud-testing tool it says was written entirely by AI agents
AWS Emulator lets developers test cloud interactions on their own machines. Automated checks govern its development, but it cannot validate an application’s access controls.
AWS has released AWS Emulator, a local substitute for cloud responses, with code it says agents developed and AWS engineers neither wrote nor reviewed. Changes pass through cloud-verified fidelity tests and automated compilation, security and regression gates before release; AWS publishes per-service match scores that can fall as testing expands. The tool supports local application and coding-agent workflows without deployment, but it does not enforce AWS permissions or model cross-account and cross-Region behavior, so it cannot replace security testing on AWS.
01
Environment variables redirect the AWS CLI, SDKs and coding agents to a local HTTPS endpoint without requiring changes to application code.
02
Kiro produced a passable first version of Amazon SQS in 15 minutes; AWS engineer Matthew Goldberg says building trustworthy guardrails took substantial engineering.
03
Rust’s embedded mode gives each test instance separate state, allowing parallel tests without sharing buckets, tables or queues.
Developers can now test AWS applications on their laptops using a tool that AWS says its engineers neither wrote nor reviewed. Released on November 11, AWS Emulator substitutes local responses for cloud calls. Its development relies on agents checking other agents’ work, with automated gates deciding which changes are allowed to ship.
The emulator runs on a developer’s computer and accepts requests through a local HTTPS address. The AWS command-line tool, software development kits and coding agents can talk to that address using the same APIs—the interfaces applications use to request cloud services. Developers set environment variables to redirect those requests; their application code stays unchanged.
That lets a test create a database table, write an item and read it back without deploying anything to an AWS Region. A coding agent can use the same loop to create resources and inspect the result locally. AWS says the emulator has no additional service fee; ordinary AWS resource charges apply when developers deploy on AWS.
Linux: install an RPM or Debian package.
macOS and Windows: build and run a Docker container using the documented Dockerfile. AWS does not publish a ready-made image.
Rust developers also get an embedded mode that runs inside the test process rather than through a separate network server. Each embedded instance keeps its own state, allowing parallel tests without sharing buckets, tables or queues.
AWS senior principal engineer Matthew Goldberg describes the project as a weekend experiment that needed substantial engineering before release. Kiro produced a passable first version of Amazon’s queue service, SQS, in 15 minutes. The larger task was building guardrails that would make continued autonomous development trustworthy, according to Goldberg’s account.
The foundation is a set of fidelity tests: short workflows that establish how closely the emulator matches real AWS. Agents devise tests from publicly released AWS API models and verify them against the cloud. Goldberg says the suite contains tens of thousands of tests and continues to grow.
Separate agent teams write those tests, find failures and fix the underlying code. Issue-finding agents group failures that may have a common cause. A fixing agent receives one issue, edits the code and submits its change to a gate checking compilation, coding standards, security and whether previously matching behavior has broken. Failed checks send the change back for rework.
Passing that gate is not the final publishing step. Changes are built and retested before entering a release branch and being packaged as a downloadable binary. Only at the end does AWS publish the code to GitHub, under a CC0 public-domain dedication.
AWS publishes per-service fidelity scores with each release, showing passed tests as a fraction of the total. Matching means matching every call’s status, response body, exact error code and message, plus the resources and data created. Scores can fall when new tests expose gaps, even while the software improves. Unimplemented operations return HTTP 501 rather than silently succeeding.
The most consequential limit concerns permissions. The emulator does not evaluate identity and resource access policies as AWS does, so it may allow a request the cloud would deny. Each instance also represents one account, without cross-account or cross-Region behavior. AWS explicitly warns against using it to test whether an application is secure.
For developers adopting the tool, AWS’s instruction is concrete: keep real credentials out of test jobs, then test security settings and the complete application on AWS before production.
Sources
aws.amazon.comTest your applications against AWS locally with AWS Emulator | Amazon Web Services
aboutamazon.comHow we built the agents that built AWS Emulator
Reader comments
Newest comments first. Replies stay oldest first.