AWS Emulator proves agents need a verification loop. Realism is the hard part.

The article looks at AWS Emulator, a laptop-sized copy of more than 200 AWS services that coding agents can test against. It starts with a story from Swami Sivasubramanian, who saw an agent's rebuild of a distributed lock service go from four nights to about two hours once the agent had a spec and a test environment to check its own work against. The author argues that an agent's verification loop is only as good as what it runs against. Mocks written by an agent tend to confirm its own assumptions, and even a faithful emulator is still a model, lacking things like throttling and regional outages. The emulator fits AWS well because AWS has a complete machine-readable spec and a real system to check against, which most teams do not have for their own services. The main gap is in internal contracts across microservices, configuration, and data that only exist in a running cluster. The author proposes that a realistic loop needs isolated, fast-starting, production-like environments, along with the team's existing integration tests. The conclusion is that emulators for managed cloud services and realistic environments for owned systems are layers in the same loop, not alternatives.