Foundry is a comprehensive training and evaluation platform designed for web-native AI agents. As AI progresses from static models to dynamic agents capable of real-world interactions, Foundry provides the necessary infrastructure to ensure these agents operate reliably within the complex and ever-changing web environment. By offering high-fidelity simulations of real websites and workflows, Foundry enables rigorous, reproducible testing of agents on end-to-end tasks under realistic conditions. Key Features and Functionality: - Deterministic Environments: Foundry offers frozen content with website versioning, allowing for consistent evaluation runs. This ensures that performance differences are due to agent behavior rather than changes in web content. - State-Based Evaluation: The platform provides structured state JSON and manages state on the backend, enabling users to define custom evaluation and reward functions based on specific criteria. - Informed Data Collection: In instances of agent failure, Foundry collects demonstration data for behavioral cloning or similar simulation examples for on-policy reinforcement learning, facilitating continuous improvement. Primary Value and Problem Solved: Foundry addresses the challenges faced by AI agents operating in web environments, such as silent failures caused by unforeseen layout shifts, DOM mutations, and race conditions. By providing a controlled and reproducible testing environment, Foundry allows researchers and developers to build reliable agents through iterative improvement. This is particularly significant given that over 60% of global knowledge work is mediated through browsers. Agents that can effectively navigate and automate web-based tasks unlock substantial economic opportunities across support, sales, and internal operations.