We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Oct 30·5 min read
How we built OWL, the new architecture behind our ChatGPT-based browser, Atlas
TL;DR
Browser integration with ChatGPT required tight coupling to Chromium, creating slow startup times and limited UI customization. OpenAI built OWL, a decoupled architecture separating the browser engine from the AI layer, enabling faster initialization and richer interfaces.
✦ Why It Matters
Engineers can decouple AI from browser engines to achieve faster startup, independent iteration, and autonomous web task automation.
Key Takeaways
- Separate browser engine from AI layer to enable independent development and faster initialization cycles.
- Design modular interfaces between Chromium and reasoning systems to reduce architectural coupling.
- Implement agentic browsing to allow ChatGPT autonomous web navigation and task execution.
- Measure startup latency and UI responsiveness as key metrics when decoupling monolithic browser-AI systems.
Full Summary
Building ChatGPT Atlas revealed architectural constraints: traditional browser-AI integration tightly coupled Chromium (the open-source browser engine) to application logic, causing slow startup times and restricting UI customization. OWL decouples these components, isolating Chromium from the ChatGPT reasoning layer through a modular interface.
This separation allows the AI system to operate independently of browser rendering, enabling faster application launch and richer user interface design. The architecture supports agentic browsing—ChatGPT can autonomously navigate, interact with, and extract information from web pages without constant user input.
Engineers can now iterate on UI and AI capabilities separately, reducing interdependencies. The decoupling also improves startup performance measurably and provides a foundation for scaling multi-agent browser automation tasks.