MangaForge

MangaForge – Agent-Generated Construction Readiness Console
MangaForge makes construction sites agent-actionable. Foreman Rita, a manga-style site foreman avatar, works with specialist agents to diagnose real-time site gaps, quantify cost leakage, and dynamically generate the exact interactive interfaces needed for immediate corrective action.
Instead of returning plain text suggestions like a traditional chatbot, the system uses a lightweight GAN-inspired proposer-critic harness (Foreman Rita as Generator + Site Inspector as Critic) to propose, evaluate, and refine rich A2UI payloads in real time. The result is a living command center where the agent literally becomes the frontend: it renders evolving dashboards, delay diagnosis cards with dollar-impact estimates, material shortage resolvers, gamified safety checklists, and one-click approval forms that users can interact with directly.
Technical Stack & Execution

A2UI (Google): The agent outputs declarative JSON that defines components (maps, cards, progress bars, interactive checklists, button groups).
CopilotKit + AG-UI: Handles bidirectional streaming so the UI updates live as the agent reasons, calls tools, and refines proposals.
MCP (Model Context Protocol): Provides standardized tools (get_site_progress, diagnose_delay, get_material_inventory, approve_pour, etc.) that ground the generated UI in real (mocked) data.
GOMS reasoning inside the orchestrator ensures the agent decomposes user goals into clear sub-goals and always produces actionable, context-aware interfaces.

This creates a powerful feedback loop: user prompt → agent tool calls → UI proposal → critic evaluation → rendered interactive surface → user action → live AG-UI update. The entire experience is impossible with chat alone — the interface itself is generated and updated at runtime by the agent.
Built in a single focused 6-hour session using the official CopilotKit starter kit, the project demonstrates strong originality through the manga avatar + GAN-style harness, clean technical execution with working interactive flows, and faithful implementation of the core hackathon theme: agents that don’t just talk — they render the exact functional UI you need in the moment.
From site gaps to site actions — generated by agents.

This project explores the edge of Agentic Intelligence, and UI