AgentOps Flight Deck

AgentOps Flight Deck is a generative UI control tower for AI agent failures. Instead of returning a plain text explanation when an agent tool call fails, the system turns the failure into an interactive operational interface: a timeline of agent steps, evidence cards, risk gates, repair options, approval queues, trace views, and final run reports.

The core idea is that serious AI agents should not be black boxes. When an agent calls tools, uses external data, or proposes risky actions, users need more than a chat response — they need a live interface to inspect what happened, understand the risk, choose a repair, and approve or block the next step.

The project uses CopilotKit as the agentic frontend layer, AG-UI-style state updates for streaming agent/application state into the UI, and a catalog-based generated UI approach inspired by A2UI principles. The agent does not render arbitrary HTML or unsafe code. Instead, it emits structured UI intents that map to trusted React components such as timelines, evidence panels, approval queues, risk gates, and reports.

Google DeepMind Gemini powers the agent reasoning layer, while LangGraph/LangChain-style orchestration is used for agent flow structure and tool-driven behavior. MCP Apps are used to expose AgentOps surfaces as reusable tool-backed UI widgets, allowing the agent to present operational dashboards and control panels instead of only text.

The demo includes a judge-ready AgentOps flow where a failed browser_search tool call is transformed into a complete interactive dashboard: timeline, evidence, medium risk gate, repair option, pending approval item, and final report. A one-click Judge Mode makes the demo deterministic and easy to evaluate, while trace import shows how external agent traces can be converted into safe generated UI.

This project directly targets the hackathon theme: moving past the chatbot bubble. AgentOps Flight Deck is not a chatbot with a nicer skin — it is an agentic interface where the agent creates the exact control surface needed at runtime so users can inspect, repair, and approve AI agent behavior.