WIGENT

Solo brainstorming is biased, team discussions take hours, and AI chat tools give single-perspective responses with no debate or convergence. There's no way to get the clash of viewpoints that produces strong ideas — quickly, with a tangible deliverable.

The Problem We Solved

Brainstorming a business idea is broken in three ways:

• •Solo thinking is biased. You fall in love with your own idea and can't see blind spots. There's no one to say "who would actually pay for this?"
• •Team brainstorming is slow. Scheduling, facilitating, keeping people on track — it takes hours to reach a conclusion that might still be half-baked.
• •AI chat is one-dimensional. ChatGPT gives you a single, agreeable response. There's no debate, no tension, no convergence from opposing viewpoints.

We wanted the collision and convergence of a real team debate — but driven by AI agents, in minutes, with a tangible deliverable at the end.

System Architecture

WIGENT is a Next.js 16 full-stack application with three cleanly separated layers:

This separation means the orchestrator knows nothing about HTTP, the API route knows nothing about debate logic, and the frontend knows nothing about GPT-4o. Each layer communicates through typed SSE events — 13 event types, all defined in a shared types.ts.

Four Pivots in Three Hours

The initial PRD described a straightforward chat UI with result cards. Over 3 hours, four pivots transformed it into something much more impactful. Each pivot followed the same pattern: problem noticed → solution designed → implemented within 30 minutes.

Pivot 2 — Rounds → Free Debate (30 Turns)

The original design used 3 structured rounds: R1 Brainstorming → R2 Debate → R3 Convergence, with 10 GPT-4o calls total. This was scrapped in favor of a 30-turn free debate with automatic phase transitions.

Why: Real meetings don't have "Round 2" announcements. The round transitions felt artificial and created edge cases in the orchestrator logic. Free debate with phase-based prompt adjustments ("argue harder" in mid-phase, "converge now" in late-phase) produced more natural conversations and far fewer bugs.

Trade-off: GPT-4o calls tripled (10 → 35+), increasing cost and debate duration from 2-3 to 5-8 minutes. We accepted this because debate quality and demo stability mattered more than speed at a hackathon.

Pivot 4 — Forced Convergence Prompts

With 30 turns of free debate, agents kept introducing new ideas right up to the last turn. The debate would end without consensus, and the summary prompt couldn't extract a coherent conclusion from a divergent conversation.

The fix was simple but critical: after turn 25, the system prompt changes to force convergence:

This transformed the last 5 turns from divergent brainstorming into natural consensus-building. The agents started agreeing, refining, and producing actionable conclusions — exactly what the summary prompt needed to generate a strong final result.

Technical Deep Dive

SSE + AsyncGenerator Pattern. The orchestrator is an async generator function that yields typed SSE events. The API route consumes these with for await...of and writes them as SSE text. The frontend hook parses the SSE stream and dispatches actions to a useReducer. This three-layer separation keeps each concern isolated — the orchestrator doesn't know about HTTP, the route doesn't know about debate logic, and the frontend doesn't know about GPT-4o.

Speaker Selection Algorithm. Each turn, the system selects the next speaker by: (1) filtering to online agents only, (2) excluding the last speaker to prevent consecutive turns, (3) sorting by speak count ascending to ensure balanced participation. Simple, but effective — every agent gets roughly equal airtime.

Agent Persona Engineering. The quality of debate is entirely determined by persona design. Generic descriptions like "market-focused thinker" produce boring conversations. Specific speech patterns produce engaging ones: the PM says "근데 이거 누가 쓰는데?" ("But who would actually use this?"), the Designer says "유저가 3초 안에 이해 못 하면 실패" ("If the user can't understand it in 3 seconds, it's a failure"). These speech habits make the agents feel like real people arguing in a meeting room.

Random Inter-Agent Delay. Between each agent's turn, a random delay of 800-2500ms is injected. Without it, agents respond instantly one after another, breaking the illusion of a real conversation. This small detail significantly improves demo immersion.

GPT-4o Refusal Handling. GPT-4o sometimes refuses to generate landing page HTML, treating it as a security concern. We addressed this by: (1) switching the landing page prompt to English, (2) framing it as "a design prototype with placeholder demo content," (3) detecting refusals by checking for <!DOCTYPE or <html tags, and (4) falling back to a minimal HTML template when generation fails.

Lessons Learned