Orchestrating Chaos

Collaborative Research Project

Digital Futures Workshop 2025

This project targets a persistent bottleneck in circular construction: valuable reclaimed materials are discarded because sorting, assessment, and reuse workflows are fragmented and hard to operationalize. We present a dialogue-driven, LLM-based multi-agent workflow that lets non-technical users co-design material-aware 3-D pieces via voice or text.

A Triage Agent interprets user intent and routes subtasks; a Catalogue Agent turns spoken descriptions of reclaimed parts into structured metadata (geometry, condition, tags); and a Design Agent combines the target form with the live catalog and drives Rhino/Grasshopper through the Model Context Protocol (MCP) for precise cataloging, packing, and shape optimization. The loop runs on smolagents’ ReAct paradigm—perception → reasoning → action → memory—yielding on-the-fly self-correction and reduced hallucinations. Compared with prior LLM pipelines (e.g., code-only graph synthesis or manufacturing-oriented checks without parametric control), our workflow couples generative reasoning with deterministic tools where each is strongest.

Next steps include integrating structural validation and fabrication-feasibility tools to ensure digitally composed assemblies can be verified and built, closing the loop from reclaimed stock to real-world construction.