YouTube Video
Project Description
COS.ai is a proactive “Chief Operating System” for everyday life: a household agent that turns reasoning into interactive, runtime-generated control surfaces instead of chat replies. When COS.ai detects a life moment, such as an evening collision between dinner, macros, workout recovery, PSV kickoff, email, tickets, and bills, it generates a task-specific UI with timelines, decision cards, charts, approval flows, reminders, and audit trails.
Users steer the agent directly through the generated interface: choosing meals, adjusting workout intensity, opening persona-specific dashboards, approving draft actions, snoozing reminders, refreshing generated views, and finalizing plans. This solves problems chatbots handle poorly because the output is spatial, stateful, glanceable, and action-oriented. COS.ai does not ask users to read a long answer and mentally translate it into tasks; it creates the operating surface for the situation.
Technically, COS.ai is built with Next.js, React, Tailwind, Recharts, Zod validation, CopilotKit, AG-UI, and an A2UI-ready architecture. A Gemini 2.5 Flash planner, routed through our agent layer and LangGraph scaffolding, receives deterministic tool data and returns a structured widget plan. The model can choose only allowlisted widget types, order, priority, copy, layout intent, and actions. Trusted React components render the final UI, so no raw HTML, CSS, arbitrary components, or unsafe actions are accepted. All UI elements came to life with Google Stitch.
The architecture is designed for robust agentic UI: fake local tools simulate household signals, Gemini plans the interface, Zod validates the plan, CopilotKit/AG-UI moves events between frontend and agent, and React renders the trusted component catalog. Every consequential action is approval-first and logged, making COS.ai a visible, steerable, and governed alternative to invisible autonomy or simple chatbot output.
Prior Work
Before the hackathon, we only prepared the development environment: installed dependencies, cached protocol/tool documentation, set up schemas, and organized local agent/project docs. No product code, COS.ai UI, persona flows, Gemini planner, fake data world, or demo interactions existed before the hackathon; those were created during the hackathon. The idea itself came to us during the initial keynote.