Captain Cool is an agentic AI system that simulates a cricket captain using Gemini-powered agents to analyze live IPL match situations and debate tactical decisions.

The agent roles unfold in a structured flow: a Stats Analyst gathers live stats and weather, a Strategist proposes a call, a Devil’s Advocate challenges, the Strategist revises, and a Match Commentator renders the final decision in fan-friendly language, with each turn revealing reasoning and the final output.

The architecture combines FastAPI backend, Next.js frontend, and Gemini AI, orchestrating five specialized agents—Stats Analyst, Strategist, Devil’s Advocate, Decision Maker (Captain Cool), and Match Commentator—to analyze cases and debate strategies before finalizing decisions.

Key learnings emphasize multi-turn reasoning, synchronous tool usage, improved UX, and memory across overs, with potential enhancements like memory caching and real-time scrapers for live states.

The project is open-source, providing full prompts, debate loop, tools integration, and run instructions, with setup requiring API keys and virtual environments.

Technical challenges included import errors, API connectivity and CORS issues, and Gemini quota limits, addressed by structured project setup, CORSMiddleware, correct API base configuration, fallbacks, prompt optimization, and lighter models for speed.

Stretch goals include real-time scraping, voice output, confidence/counterfactuals, and memory integration, using Google Gemini ADK and live data sources; links point to GitHub, prompts, and a demo.

The system performs real-time tool calls to fetch live context like weather, rather than relying on hardcoded data.

Observability and telemetry expose orchestration cycles, token throughput, active agents, memory retrieval, and tool traces to ensure explainability and liveliness.

Lessons highlight the value of precise system prompts, dissent from the Devil’s Advocate, and the impact of real API data on output quality, with dew factor often decisive.

The main engineering challenge was balancing believable tactical intelligence with orchestration complexity, UI responsiveness, and explainability, addressed by visible reasoning evolution and structured orchestration.

Stretch goals include explainability, memory across overs, and confidence/counterfactuals, aided by Streamlit session state to retain strategy in match history.