How I orchestrate 19 specialized AI agents across 86 data sources to generate sales intelligence reports (~97 LLM calls per report)

Built a multi-agent system called Recon that generates deep prospect research reports. Thought the architecture might be interesting to this community. Agent specializations: Each of the 19 agents owns a specific research domain - competitive intelligence, financial forensics, buying committee mapping, psychological profiling (13 dimensions), engagement strategy, and more. They don't just retrieve data - they analyze, cross-reference, and synthesize. Data sources: 86 different sources including SEC filings, earnings call transcripts, patent databases, job postings (hiring signals), tech stack detectors, social signals, news sentiment, and more. Orchestration: The brain (orchestrator) coordinates which agents to deploy based on what's known about the target company and what intelligence gaps exist. A single report runs ~97 LLM calls total. Each specialist agent operates independently but the orchestrator synthesizes their outputs into coherent intelligence. Key challenges we solved: - Agent coordination without bottlenecking (parallel execution with dependency awareness) - Source attribution (every claim traceable to which of the 86 sources produced it) - Conflicting signal resolution (when sources disagree) - Balancing depth vs speed (full reports in minutes, not hours) - Preventing hallucination in synthesized output Stack: Electron + React frontend, Fastify + Prisma backend, BullMQ for job orchestration. Happy to go deeper on any part of the architecture. If you want to see the output, DM me.