A2Atrium

A2Atrium is a runtime contact surface that assembles interactive UI around the specific moment of meeting someone — not a CRM, not a chatbot, not a static page. When you encounter someone at an event, three distinct intent-driven interfaces are generated on demand: Follow up (draft message, time slots, shared context bridge), Convene (micro knowledge graph of conceptual overlap, collaboration proposal, vesica portal), and Remember (annotation field, encounter history). The interface IS the answer — not a reply that tells you what to do, but a surface that has already done it.

How it moves beyond chat: The conversational layer is Telegram — you text your agent a LinkedIn screenshot with a caption note; Gemini Vision extracts the contact, your caption becomes the encounter note, and the canvas surface appears within 15 seconds. No UI to open, no form to fill. The rendered surface is assembled by an agent emitting a JSON spec against a typed component catalog; the renderer resolves the spec into live, interactive components. Three different intents produce three structurally different interfaces, all anchored to the same persistent contact edge. The agent cannot invent UI — it assembles from trusted primitives. That catalog is the guardrail.

Stack: CopilotKit (useFrontendTool({render})) for controlled-tier component registration; A2UI declarative spec pattern for intent-driven surface assembly; LangChain Deep Agents for orchestration; Gemini 2.5 Flash (vision + generation); AG-UI for agent ↔ frontend state sync; python-telegram-bot v22 for the conversational input layer; Next.js 15 frontend; MCP Apps via mcp-use for a third rendering surface (architected). Place-aware theming stamps each surface with the visual palette of the event where the encounter happened. The follow-up agent runs a two-step “think first” prompt — gold extraction before drafting — ensuring the surface reflects the actual encounter, not generic warmth.

Originality: A vesica — borrowed from the geometric figure where two circles overlap — is the navigable entry point between two latent spaces. The Convene surface generates this: a proposed shared workspace shaped by the conceptual overlap between two people, with a portal that, in production, spins up a shared A2UI surface accessible to both parties’ agents. This is the prototype move toward spatial multi-agentic observability: making harmony between agents visible and giving humans meaningful choice over how their agents cooperate.