Multi-Agent Research Pipeline — Planner + Workers + Critic

Key Takeaways

• The Planner decomposes the research objective into bounded subtasks with explicit JSON schema contracts — each Worker receives a scoped prompt with only the context it needs, preventing context-window saturation across the fleet.
• Worker agents run in parallel fan-out (analogous to async microservice calls behind an API gateway), each returning a structured result object; the Orchestrator aggregates only after all Workers reach terminal state or a defined timeout threshold.
• The Critic agent performs contradiction-detection and confidence-scoring across aggregated Worker outputs before synthesis — this reflection loop is what separates a production-grade pipeline from a prototype that fails silent-failure audits.
• Memory anchor — API-gateway-microservices: Planner is the gateway, Workers are the services, Critic is the schema-validation middleware; route only scoped context, validate before aggregation, never let raw Worker output hit the consumer directly.