ABRA is a research prototype that replaces the chat window as the main surface for working with AI agents. Instead of scrolling through a conversation to find a result, each agent owns a persistent panel, called a view, that it can update as it works. The panel stays in place across sessions and can be shared with other agents or team members. You see the current state of the work at a glance, without digging through message history. The system treats UI panels as first-class outputs, not decoration. An agent writes structured data to its panel, and the panel renders that data as interactive controls. A second agent can read the same panel, act on it, and write its own output back. The research team tested the approach with a working prototype and scored agent outputs using a language model as the judge. Results showed that structured, panel-based outputs were easier to assess and act on than free-form chat replies. For designers, the central tension is trust. You have to decide whether a panel built by an agent is safe to interact with, and whether the controls in it do what they appear to do. ABRA surfaces this by making the agent's authorship of the panel visible. The pattern is close to generative UI, meaning the agent builds the interface as it works, but adds a binding between the agent's identity and the surface it produces. You can see which agent made which panel.