World Model for Browser Agents
Infrastructure to ship real apps with AI - end-to-end in the browser
We’re building the an end-to-end web simulator for browser agent training and evals. Our system enables teams to test, benchmark, and optimize browser automation models at scale. - Deterministic Web Simulation → Stable, reproducible testing with versioned web snapshots. - Live Web Evaluation → Identify failures caused by UI drift, captchas, and dynamic content. - Automated Annotation & Labeling → Generate high-quality training data for benchmarking. - RL-Driven Agent Optimization → Improve models with scalable, feedback-driven learning. By combining synthetic user simulations, automated evaluations, and large-scale benchmarking, we help teams build more reliable web agents that handle real-world environments with confidence.
We’re building browser-native infrastructure that makes software creation production-ready end-to-end. Today’s AI tools can scaffold code, but they fail on the critical workflows that actually get products into production — provisioning, auth, QA, and compliance — all of which happen inside the browser. Our system makes these workflows deterministic, reproducible, and automatable, so both agents and humans can go from idea to a production-ready app. The opportunity is massive: billions of dollars of value locked up in workflows that today block non-technical founders and slow technical teams. We’re building the infrastructure layer that makes production possible.
Foundry is shifting from a tool focused on simulating and evaluating browser agents (mainly for AI training and benchmarking) to a broader browser-native infrastructure that automates the critical steps of shipping production-ready apps, targeting real-world workflows involving both agents and humans. While related to automation and browsers, this is a significantly different and more expansive product direction.