Multi-agent reasoning enables predictive design of living materials

Abstract

Artificial intelligence is increasingly used to accelerate scientific discovery, but most successful frameworks operate within well-defined molecular, protein or materials spaces. Living materials present a more formidable computational problem because functions emerge from context dependent coupling among cells, matrices, fabrication processes and evaluation conditions. Here we introduce LiveMat, a multi-agent reasoning framework that transforms unstructured literature into a computable design space for living materials. LiveMat standardizes 34,215 living material records, integrating 16,769 microorganism and 17,446 polymer entries into a knowledge graph linking living components, abiotic matrices, functional outputs, evaluation contexts and performance metrics. Benchmarking across five large language models shows that living material reasoning is limited mainly by cross-domain feature integration rather than coarse classification. LiveMat overcomes this limitation through constraint decomposition, provenance-aware extraction, consistency checking and expert-anchored ranking. In a prospective wound-healing task, it prioritizes a four-component design with state-of-the-art in vivo performance, establishing a scalable infrastructure for interpretable, evidence-grounded living material discovery.