Inspiration
Most “meeting AI” tools are passive: they summarize after the conversation is over. We wanted to build something fundamentally different—an AI that participates during the discussion, challenges bad assumptions, and actively improves the quality of thinking in real time.
What it does
Agentic Think Tank is an opinionated AI participant that joins calls or live conversations. It listens continuously, interrupts when claims are incorrect or weakly reasoned, injects verified facts, and spins up background research agents to surface relevant articles, data, and media mid-discussion.
How we built it
We built a real-time system using a React frontend and a Node.js backend connected via WebSockets. The browser handles live speech-to-text, while the backend maintains conversation state, applies intervention logic, and orchestrates agent responses. Voice output is generated using text-to-speech so the Oracle can speak back naturally.
Challenges we ran into
Real-time systems are unforgiving. Managing low-latency voice transcription, deciding when to interrupt without being disruptive, and keeping the Oracle opinionated without being noisy required careful heuristics and iteration. Tooling constraints and runtime compatibility also forced deliberate architectural choices.
Accomplishments that we're proud of
We built a working, end-to-end agent that listens live, reasons about conversations, and speaks back in real time. Most importantly, the Oracle behaves like a participant—not a chatbot—demonstrating a new interaction pattern for AI in group settings.
What we learned
The hardest part of agentic systems isn’t model capability—it’s behavior design. Deciding when to act, when to stay silent, and how to challenge humans constructively is far more important than raw intelligence.
What's next for Agentic Think Tank
Next, we want to add richer research agents, source credibility scoring, tighter interruption controls, and deeper integrations with platforms like Discord and Zoom—turning the Oracle into a trusted, always-on thinking partner for teams.
Log in or sign up for Devpost to join the conversation.