.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
AetherFlow: Autonomous Material and Business Agent
Band of Agents is a fully autonomous multi-agent system that automates the entire materials R&D workflow. A user provides a single natural-language material requirement (e.g., "I need a heat-resistant ceramic for turbine blades"), and three specialized AI agents collaborate autonomously — with zero further human prompting — to deliver a complete executive product proposal. Agent 1 (Material Theory) uses LangGraph + Groq (llama-3.1-8b) to convert the requirement into a validated material specification through a 5-node pipeline (parser → planner → reasoner → generator → checker) with a self-correction loop. It grounds outputs using real APIs — PubChem for compound verification, Wikipedia for summaries, and a hardcoded periodic table for oxidation state and cost checks. Agent 2 (Virtual Lab) uses CrewAI + GPT-4o to simulate synthesis via two roles: a Process Simulation Scientist (feasibility, yield, conditions) and a Manufacturing Cost Analyst ($/kg breakdown, GO/NO_GO verdict). Agent 3 (Procurement & Sourcing) uses pydantic-ai + GPT-4o to perform supply-chain lookups and deterministic ROI computation, producing a final ExecutiveProductProposal with a GREENLIGHT/PILOT/HOLD/REJECT recommendation. All agents communicate over Band's WebSocket platform, passing structured Pydantic JSON schemas autonomously. Key design choices include: a self-correction loop with configurable retries, real API grounding to prevent hallucination, deterministic financial/cost tools so LLMs never fabricate numbers, and subprocess isolation for nested agent runtimes. If Agent 2 rejects a material, Agent 1 retries up to 3 times before reporting failure. Each agent uses the best-fit AI framework for its task — LangGraph for stateful graph pipelines, CrewAI for multi-role simulation, and pydantic-ai for structured proposal generation.
19 Jun 2026
