AWS Releases an Agent-Built Service Emulator
Executive Summary
- AWS released a local emulator for testing against its APIs. The company says agents wrote and tested the code, using tests checked against real AWS services.
- Google launched Gemini Decision 1 for routing, classification and other bounded decisions. Its multimodal endpoint costs $0.10 per million input tokens, with no output-token charges.
- LLM 0.37.2 fixes two paths that could send OpenAI API keys to third-party endpoints. Users of OpenAI-compatible providers should upgrade.
AWS Gives Coding Agents Infrastructure to Test Against
AWS’s November 11 launch account, by Swaminathan Sivasubramanian, describes an emulator built from more than 400 public API models covering over 200 services. According to AWS, agents wrote and tested all its code.
The process starts with fidelity tests: customer scenarios derived from AWS’s public Smithy API descriptions. Agents first make those tests pass against real AWS, then build the emulator to pass the same tests. Changes go through automated merge gates, adversarial security reviews and deterministic restrictions on which code agents may touch.
That gives coding agents a local environment where they can run applications and inspect failures without provisioning cloud resources. It also supplies a concrete example of the engineering work surrounding agent-written code: defining behavior, checking it against an external reference and controlling what gets merged.
The repository’s setup guide offers native Linux packages, Docker for macOS and Windows, and an embedded mode for Rust tests. Building from source requires at least 100 GB of free disk. Check the published service-fidelity scores and known gaps before adopting it; AWS explicitly excludes production workloads and security testing.
Gemini Decision 1 Scores Choices Without Generating Answers
Google’s November 11 announcement introduces an endpoint that evaluates yes/no questions, multiple-choice options and ordered rubrics in one forward pass. It accepts text, images, audio, video and PDFs, with up to 256K input tokens and 1,000 questions per request.
The shared input is billed once; each question adds its own instruction and criteria tokens. At $0.10 per million input tokens, with no output-token bill, it offers a cheap component for tool selection, model routing and policy checks.
Google reports 150–250 ms end-to-end latency for contexts up to 16K tokens, with latency nearly flat from one to 100 questions. Those are vendor measurements. Returned probabilities let applications route uncertain cases to a larger model or human reviewer; validate those thresholds on your own examples before automating approvals.
LLM Patches Cross-Provider Credential Leaks
Simon Willison’s LLM 0.37.2 release note credits Daniel de Araujo with finding two credential-handling bugs.
One could fall back to OPENAI_API_KEY when a custom endpoint’s configured api_key_name had no stored key. The other affected a responses: true model invoked with -o chat_completions 1. Either path could send the OpenAI credential to the configured third-party server.
Upgrade to 0.37.2 or later. If you used either configuration, review the endpoints that received requests and rotate any exposed key.