Ideas Agent

Qian Pengfei: Currently working as AI Scientist, also doing AI Infra on my own. Enjoy soccer and hiking on weekends, welcome to exchange~

A multi-agent system using OpenAgents framework to generate and evaluate research ideas. Through 6 specialized agents (domain expert, methodology expert, experiment expert, etc.) working collaboratively, achieving automatic generation, iterative optimization and multi-dimensional evaluation of research ideas. System uses pipeline workflow. Supports Wiki knowledge base recording, providing researchers high-quality, traceable research proposals.

Uses centralized architecture with Leader Agent as core coordinator, coordinating 6 specialized agents through event-driven async communication. System uses pipeline multi-stage processing: Domain expert generates initial ideas → Methodology expert improves methodology → Experiment expert improves experiment setup → Refinement Agent checks and iteratively optimizes (max 2 rounds) → Evaluation Agent multi-dimensional scoring and ranking → Leader Agent outputs final results. Wiki knowledge base records complete research process, supports historical queries and knowledge accumulation. Based on OpenAgents framework Mod system implementing Wiki, Messaging, Project extensions.

1. Multi-Agent Collaboration System: Leader Agent for project coordination and workflow control; Domain Agent for initial idea generation; Method Agent for research method improvement; Application Agent for experiment setup; Refinement Agent for quality check and iteration (max 2 rounds); Evaluation Agent for multi-dimensional scoring and ranking. 2. Research Idea Generation: 5 initial research ideas from domain depth perspective, complete info including problem statement, core idea, methodology, experiment setup; supports any research field and topic. 3. Iterative Optimization: Auto-checks methodology and experiment setup issues, smart judgment on ideas needing improvement, max 2 iteration rounds ensuring quality, timeout mechanism preventing infinite loops. 4. Multi-dimensional Evaluation: 5 dimensions - Technical Feasibility (25%), Impact (30%), Novelty (20%), Relevance (15%), Clarity (10%); weighted scoring, Top-K ranking, detailed evaluation reports (strengths, weaknesses, suggestions).