composable simulator infrastructure

A composable simulator infrastructure is a toolkit you can mix and match like Lego blocks to build simulations for different scenarios. Instead of writing a whole new simulator from scratch for each AI system you want to test, you use pre-built, interchangeable components that snap together. This lets you quickly test new features, like a robot that works across your phone and laptop, without rebuilding everything, making the simulators faster to create and easier to update as the technology changes.

The papers all build simulators using this modular approach: one creates a framework to safely test AI agents by combining risk taxonomies and executable test cases, another simulates agents coordinating across different operating systems, and a third uses a coding agent to automatically extend a simulator's components. They find that without composable design, simulators are slow to update and miss critical failures, like agents failing when they must juggle tasks across multiple devices or handle advanced attacks. By making infrastructure composable, these tools reveal dangerous gaps that simpler, fixed simulators overlook.