Amazon Web Services launched AWS Emulator on Wednesday, a free, open-source local testing suite built entirely by autonomous AI agents to replicate its cloud services on developer machines. The tool allows programmers and coding bots to run multi-service cloud workflows on Linux, macOS and Windows without incurring cloud infrastructure fees.
The software replicates the application programming interfaces of AWS services locally, as InfoWorld reported. Developers redirect standard command-line tools and software development kits by pointing the AWS_ENDPOINT_URL environment variable to their local machine, running existing test suites offline without modifying application code.
Replicating the cloud on a laptop
The tool addresses a growing problem created by the rapid adoption of AI coding assistants such as AWS Kiro and Claude Code. While these automated assistants generate code at high speed, validating that code against remote cloud infrastructure creates delays and inflates cloud bills.
Advait Patel, a senior site reliability engineer at Broadcom, said agentic workflows require tight feedback loops that become expensive when executed directly against cloud accounts. “Moving that loop to a laptop cuts each cycle from minutes to seconds,” Patel said.
Local execution also prevents test environments from cluttering corporate accounts with abandoned resources. Ashish Chaturvedi, executive research leader at HFS Research, said developers can test complex workflows across services such as S3, Lambda, DynamoDB and API Gateway without leaving behind idle virtual machines or storage buckets that accumulate charges months later.
To keep system overhead low, the emulator was written in the Rust programming language. It operates as a local HTTP service or within a Docker container, supporting all 426 AWS cloud services at launch.
Three teams of autonomous AI agents
The creation of the tool marks an unusual milestone in automated software development. Matt Goldberg, senior principal engineer at AWS, wrote in an Amazon blog post that autonomous AI agents designed and wrote the entire codebase, with no human engineer writing or reviewing a single line of code, as SiliconANGLE reported.
Goldberg took inspiration from Alf, a foundational Paxos-based locking mechanism built two decades ago to support Amazon S3 and DynamoDB. Swami Sivasubramanian, vice president of agentic AI and emerging technologies at AWS, explained that Alf originally required months of manual Java virtual machine simulation to test concurrency bugs and inject faults.
To replicate that testing environment at modern cloud scale, Goldberg tested whether AI agents could build replicas directly from published interface models. After an initial test in which AWS Kiro generated a working replica of the Amazon Simple Queue Service within 15 minutes, AWS expanded the setup into a multi-agent system running on Amazon Bedrock AgentCore.
The project organised dozens of autonomous agents into three separate functional groups:
- The first team synthesised thousands of end-to-end fidelity test suites from AWS Smithy API models, verifying them against the live AWS cloud as ground truth.
- The second team executed those tests against the local emulator to uncover behavioural discrepancies and automatically logged bug tickets on GitHub.
- The third team wrote, tested and submitted the software fixes to resolve those discrepancies.
Before any generated code entered the repository, an automated merge gate run by adversarial security agents checked the pull requests for vulnerabilities. AWS has published the project under a Creative Commons CC0 public-domain dedication licence on GitHub, alongside its unedited commit history and a public backlog of known discrepancies.
Fidelity gaps and production risks
Despite covering 426 services, the software contains functional gaps that prevent it from completely replacing live cloud testing. Devin Pratt, research director at IDC, pointed out that the simulation lacks critical operational constraints.
“At launch, the emulator does not simulate throttling or outages, snapshot and restore, or per-Region availability, which matter most for resilience,” Pratt said.
Patel said that missing operational behaviours such as rate-limiting errors and regional latency differences represent the exact failure modes that take systems down in live production. When local tests never encounter an HTTP 429 rate limit or an unexpected timeout, developers cannot verify whether their retry and backoff logic actually functions.
AWS measures the accuracy of the software by comparing local API responses to live cloud outputs. When the emulator encounters an operation that has not been implemented or verified, it returns an HTTP 503 error rather than returning an unverified success response.
Security risks also remain for organisations that do not manage local credentials carefully. Because the emulator does not run inside an isolated sandbox, functions that contain hardcoded production credentials can still send live calls to the actual AWS cloud during local testing.
Changing roles for human developers
For corporate IT leaders, the release offers a method to deploy AI coding agents without giving them direct access to live cloud infrastructure. Chaturvedi said that providing a contained local environment allows enterprises to let agents experiment freely without risking accidental spending or cloud security exposures.
Sivasubramanian said the project reflects a shift in daily engineering duties. Rather than writing, testing and debugging code by hand, engineers act as systems orchestrators who define test boundaries, build architectural specifications and configure guardrails for autonomous agents.
AWS said the autonomous agent pipeline will continue to update the emulator as new services launch and existing API definitions change.