Adoption and Infrastructure

• OpenAI internal adoption shifted when role-specific context became available through local file access, connectors, and computer use, rather than through any change in the technicality of the roles themselves.
• Computer use, once a "janky" developer tool, has reached a tipping point where it serves as a universal connector for tasks that lack APIs, often operating silently in the background without user intervention.6:36
• Teams are now creating reusable internal products and apps instead of just generating one-off documents, a shift that allows for rapid iteration and faster tool-building than traditional software cycles.15:30

Interface and Interaction

• Voice input is the most efficient channel for supplying context, yet users require visual outputs and mid-run steering capabilities because reading remains faster than listening and workflows are increasingly bidirectional.
• Mid-run interruption is an emerging necessity; agents require interfaces that allow users to catch errors, re-route tasks, and validate outputs in real-time to avoid wasting tokens on incorrect paths.33:33
• Proactive reporting—AI identifying blind spots without being prompted—is the next development frontier, though it requires strict confidence controls to prevent costly, noisy, or low-signal overnight inference.27:44

Concrete Breakthroughs

• The capability threshold for creative generation has crossed into physical-world coordination: agents can now generate 1,000-piece Lego sets and print customized children’s books.41:46
• Automation has moved beyond simple tasks to solving high-stakes problems, with one internal example showing an agent identifying a significant tax calculation error that saved thousands of dollars.0:35