Agent work surfaces are converging around four missing pieces: a browser, authenticated sessions, triggers, and temporary or durable memory.
Claude Cowork's built-in browser and ChatGPT's browser and automation surfaces show the same product pressure from different directions. The agent must see a working context, wait for an event or schedule, carry enough state to continue, and cross sensitive boundaries only with visible user control.
The useful architecture is not one universal browser agent. It is a set of bounded surfaces with distinct identities, permissions, recovery paths, and evidence. Teams should test the complete action loop and keep personal memory separate from shared operational state.