Multi-Agent Research System

This system explores multi-agent orchestration patterns for open-ended research questions. A query is dispatched concurrently to six configurable personas — Educator, Skeptic, Scientific, Logical, General, and Economic — each with its own system prompt, model, and grading weights defined declaratively in YAML. An orchestrator coordinates the run: it gathers agent responses via asyncio, scores them along multiple dimensions, analyzes cross-agent consensus, and synthesizes the results into a single combined write-up. The system ships with two entry points — a CLI and a FastAPI REST API — plus a Next.js dashboard for exploring runs interactively, so the same orchestration core is reachable from a terminal, an HTTP client, or a browser. Any agent can run against OpenAI, Anthropic, or a local Ollama model, making it possible to test the same persona set fully offline. Structured JSON logging, token/latency/cost tracking, and session tracing are built in, with an eye toward treating agent output quality as something to measure and compare rather than take on faith.