AppWorld: An AI Framework for Consistent Execution Environment and Benchmark for Interactive Coding for API-Based Tasks

Researchers have introduced the AppWorld Engine, a robust execution environment with 60K lines of code, featuring nine apps operable through 457 APIs to simulate realistic digital tasks for autonomous agents. The AppWorld Benchmark includes 750 diverse and complex tasks requiring rich, interactive code generation and thorough programmatic evaluation. The framework’s modularity and extensibility allow for user interface control, coordination among multiple agents, and examination of privacy and safety issues in digital assistants.