Critic-Generator Research Agent

Two-agent loop where critic refines search queries and generator synthesizes — disentangled per AgentDisCo pattern

Difficulty: weekend | Stack: Python, LangGraph, Anthropic Claude API, Tavily Search API, SQLite

Who this is for

Researchers and analysts who need coherent long-form synthesis from web search without context-window drift

Build steps

1. Define two agent nodes in LangGraph: CriticAgent (receives outline + search history, outputs refined queries + gap analysis) and GeneratorAgent (receives queries, retrieves via Tavily, returns chunks + partial synthesis)
2. Wire adversarial loop: Generator runs retrieval → Critic scores relevance and rewrites queries → repeat N rounds or until Critic confidence threshold met
3. Store each round’s queries, retrieved chunks, and critic scores in SQLite so the loop is inspectable and restartable
4. Add a final Synthesizer pass that merges all approved chunks into structured markdown report
5. Build minimal CLI: research 'topic' --rounds 3 --output report.md

Risks

• Critic and Generator using same base model may converge on same blind spots — use different temperature or system prompts to force divergence
• Tavily costs spike fast with many rounds; cap rounds and chunk deduplication is essential
• LangGraph state graph gets complex quickly — keep critic/generator state schemas flat or serialization bugs appear