<p>The water treatment industry urgently requires innovative materials to address persistent and emerging contaminants. However, conventional materials discovery processes remain slow and largely serendipitous, hindered by the ‘dark matter’ within complex purification mechanisms and stringent electronic structure requirements. Here we present ECOMATS, a multi-agent artificial intelligence system for water purification applications, which integrates expert-validated knowledge graphs with seven fine-tuned large language models to autonomously design peroxymonosulfate-activating catalysts for advanced oxidation processes. ECOMATS uses a five-dimensional evaluation framework and a triple-agent blind review system with consistency-based score fusion to improve prediction reliability. Using this integrated approach, we identified several promising catalyst candidates, with theoretical calculations showing that their <i>d</i>-band centres fall within the optimal range for peroxymonosulfate activation. A representative catalyst, (FeTCPP)Co₂(MeIm)₂, where TCPP is tetrakis(4-carboxyphenyl)porphyrin and MeIm is 2-methylimidazole, was synthesized and showed exceptional performance, achieving 90.5% degradation of perfluorooctanoic acid within 5 min, surpassing most reported analogues. It also maintained robust activity across a broad pH range of 3–11 and consistently removed contaminants in practical wastewater samples from 31 provinces across China, demonstrating practical effectiveness. This work establishes a paradigm of accelerated and application-oriented material discovery, advancing sustainable water purification.</p>