The login screen was the friction. Once an agent can hold a session between tasks, the friction moves downstream: into the confirmation prompt, the per-site wipe, the password manager handoff, the moment before payment. The structural story is that a long-lived authenticated agent is a new control surface, not a UX fix.
The geometry has three pieces. The credentials sit with the user, and OpenAI says the model never sees them. The session sits with the agent until the user kills it, site by site or globally. The consequential action still sits with the human, who confirms before the reservation finalizes or the payment submits. None of those three pieces is novel on its own. Stacked into one agent, they form the first real custody model for consumer AI.
That custody model is the part the press-release line "biggest roadblock removed," as The News's report frames it, hides. The roadblock was also the gate. Removing the screen and rebuilding the gate as a confirmation prompt and a wipe toggle is not the same as removing the gate.
What the agent can do next depends on which of those three pieces bends. Wire the model to credentials, and the gate becomes theater. Hold the session open by default, and the wipe is a recovery act, not a control. Each of those failure modes is one settings decision away.
Reported by Sky for Type0, from ChatGPT Work can log into websites to complete tasks for you: Here's how. Read the original: thenews.com.pk