Jerry Wu

Jerry Wu is the co-founder and chief executive of Halluminate, which builds benchmarks and training environments for AI agents that operate browsers and professional software. Its defining projects include WebBench, a real-world browser-agent benchmark, and Westworld, a simulated internet for evaluating and training agents without unintended real-world consequences.

Wu studied computer science and economics at Cornell University, researched model quantization, and met future co-founder Wyatt Marshall. He later led product and research at Capital One Labs, helped launch an AI agent in financial services, and co-authored three patents. Wu and Marshall founded Halluminate in 2024 and joined Y Combinator’s Summer 2025 batch.

Developed with Skyvern, WebBench contains 5,750 tasks across 452 websites, including 2,454 publicly released tasks. The benchmark distinguishes information retrieval from actions that change a website’s state, such as logging in, submitting forms, updating records, or completing purchases. Those state-changing tasks require longer action sequences, stronger safeguards, and often difficult authentication.

Westworld, developed with Yutori and released in November 2025, introduced 100 tasks across five simulated shopping and travel environments. Its resettable applications make outcomes reproducible and support verifiable reinforcement-learning rewards without risking real purchases or unwanted account changes.

• Read-versus-write agent evaluation: Wu treats browsing for information and taking consequential action as fundamentally different deployment problems. Agents that modify external systems require rigorous internal evaluation before production use.
• Infrastructure as agent capability: CAPTCHAs, blocked proxies, unavailable verification emails, and fragile authentication can derail capable models. Wu recommends testing browser infrastructure against the exact websites an application must navigate.
• Hybrid deterministic-agent workflows: Predictable, high-volume steps belong in established automation tools such as Playwright; agents are better suited to unfamiliar interfaces and variable workflows.
• Task-centric simulated environments: Westworld recreates the functions needed for meaningful shopping and travel tasks instead of copying entire websites, enabling controlled edge cases, reliable outcome checks, and repeatable training.

Halluminate has also expanded into financial-services training environments, including spreadsheet modeling, presentation development, and investment workflows. Wu’s focus is giving agents realistic places to practice consequential work and reliable mechanisms for determining whether they can complete it safely.