Agentic generative dashboards

A dashboard is usually a fixed screen of charts. The new pitch is software that builds the screen after you ask for it, with an agent choosing widgets, data, and layout for a specific job. (developers.openai.com) OpenAI’s Agents SDK says agents are applications that “plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work.” OpenAI’s platform page also says developers can build agents with Agent Builder, the Agents SDK, and ChatKit for front-end experiences. (developers.openai.com) (openai.com) That helps explain why “agentic dashboards” are showing up in demos and code. GitHub projects now describe dashboards where an AI agent listens to a prompt, selects prebuilt components, and assembles a view instead of sending back only text. (github.com 1) (github.com 2) The design pattern behind those demos is often called generative user interface, or generative UI. CopilotKit’s documentation splits it into “controlled” versions, where the developer prebuilds components and the agent decides when to show them, and more open-ended versions that render richer interface elements on the fly. (github.com 1) (github.com 2) In practice, that means the same product can look different for a sales manager, a support lead, or a finance analyst. OpenAI’s platform page says ChatKit is for “customizable, front-end agentic experiences,” and Microsoft has already shipped an Agent Dashboard aimed at leaders and analysts tracking usage and performance across organizations. (openai.com) (techcommunity.microsoft.com) The conversation-first part is central to the shift. Instead of browsing menus and filters, users type or speak a request, and the agent can fetch data, call tools, and return charts, forms, or controls tied to that request. (openai.com) (github.com) Developers are also trying to keep those interfaces constrained enough for production use. CopilotKit’s generative UI examples emphasize pre-approved components and app-owned styling, while assistant-ui’s examples show “persistent UI components” that both the user and the assistant can update. (github.com 1) (github.com 2) That caution reflects a basic product problem: companies do not want an agent inventing arbitrary controls in a business app. OpenAI’s March 11, 2025 launch post for its agent tools highlighted orchestration, tracing, evaluations, and guardrails as the infrastructure needed to turn model behavior into production software. (openai.com) The result is less a replacement for every dashboard than a new layer on top of existing software. If the current wave holds, more apps will keep their underlying data models and permissions, but let an agent assemble the screen each user needs at the moment they ask. (developers.openai.com) (openai.com)