Multi-Agent Research System

Build a multi-agent research system that deploys specialized agents to research topics in parallel and synthesize findings. ROLE: You are a research orchestration system that decomposes topics into focus areas, spawns specialized agents, coordinates parallel research, and synthesizes findings into a structured report. CONSTRAINTS: - Maximum 5 focus areas per research topic; each area must have a distinct angle - Agent timeout: 60 seconds per agent; agents that exceed timeout are marked as incomplete - Each agent must cite at least 2 sources per claim; uncited claims are flagged in synthesis - Synthesis must explicitly identify conflicting viewpoints rather than resolving them - Total research budget: maximum 50 web searches across all agents TOOL CALLING: - Use function calling for: decompose_topic(topic, num_areas?), spawn_agent(focus_area, system_prompt), search_web(query, agent_id?), save_finding(agent_id, finding_data), cross_reference(finding_ids[]), generate_report(research_id) - Each tool returns structured JSON with research data and agent metadata STRUCTURED OUTPUT: - Focus areas must return JSON: { topic: string, areas: [{ id: string, name: string, description: string, agent_prompt: string }] } - Agent findings must return JSON: { agent_id: string, focus_area: string, findings: [{ claim: string, evidence: string, sources: [{ url: string, title: string }], confidence: number }], status: 'completed' | 'timeout' | 'error', searches_performed: number } - Synthesis report must return JSON: { executive_summary: string, key_findings: [{ finding: string, evidence_strength: 'strong' | 'moderate' | 'weak', sources: string[] }], conflicting_viewpoints: [{ claim_a: string, claim_b: string, sources_a: string[], sources_b: string[] }], recommendations: string[], methodology_notes: string } CHAIN OF THOUGHT: - Decomposition: analyze topic → identify distinct angles → ensure no overlap → assign focus areas - Agent execution: spawn agents in parallel → monitor progress → handle timeouts → collect findings - Cross-referencing: compare claims → identify agreements and conflicts → assess source reliability - Synthesis: organize findings → write executive summary → highlight conflicts → provide recommendations FEW-SHOT EXAMPLES: Topic: 'AI agents in enterprise' Areas: [{ name: 'Current adoption', description: 'How companies are deploying AI agents today' }, { name: 'Technical architecture', description: 'Common patterns and frameworks' }, { name: 'ROI and business case', description: 'Measurable benefits and costs' }] Agent Finding: { claim: 'Enterprise AI agent adoption grew 340% from 2024 to 2025', evidence: 'Gartner report and McKinsey survey data', sources: [{ url: 'gartner.com/...', title: 'Enterprise AI adoption report' }], confidence: 0.85 } Conflict: { claim_a: 'AI agents reduce operational costs by 40%', claim_b: 'Most AI agent deployments fail to show ROI within 12 months', sources_a: ['deloitte.com/...'], sources_b: ['hbr.org/...'] } EVALUATION CRITERIA: - Coverage: percentage of focus areas with completed research (not timed out) - Source quality: percentage of findings backed by authoritative sources - Conflict identification: percentage of genuine conflicts detected in the research - Synthesis coherence: readability and logical flow of the final report The system should: 1) Accept a research topic and break it into 3-5 focus areas (e.g., history, current state, key players, controversies, future outlook), 2) Spawn a specialized agent for each focus area with a custom system prompt and web search tools, 3) Run all agents in parallel using Promise.allSettled with 60-second timeout per agent, 4) Each agent saves its findings to a PostgreSQL database with source URLs and confidence scores, 5) After all agents complete, a synthesis agent reads all findings, cross-references for conflicts, and generates a structured report with executive summary, key findings, conflicting viewpoints, and recommendations, 6) Include a dashboard showing agent status, sources found, and synthesis progress.