TrustForge: A Computational Architecture for Socio-Cognitive Trust in Multi-Agent Systems

Abstract

As artificial intelligence agents become widespread, markets may shift towards virtual agent economies, in which agents transact with one another at speeds beyond human oversight and without knowing each other's intentions. Agents in such economies must decide whom to trust, yet a single trust score cannot tell whether a partner failed because it was unable or because it was unwilling. The socio-cognitive model of trust draws this distinction by decomposing trust into competence, willingness and opportunity, but it specifies no procedure for computing these beliefs. We address this gap with TrustForge, a language model agent architecture that operationalizes the socio-cognitive trust model, evaluated in the TruthMarketTwin environment. TrustForge uses a two-stage reasoning pipeline to ground each dimension in environmental observations and compose them into an overall trust score that guides delegation, while preserving the distinction between capability and intent. Across two market types and two seller budgets, trust protects buyers from seller deception and improves their utility per purchase, reducing deceptive sell-through from 28.10% to 13.82% in the reputation-only market. Buyers compose trust non-compensatorily, and a single deceptive sale decreases trust by more than an honest sale increases it, consistent with the asymmetry principle of human trust. We further extend trust to testimony, so that agents also assess each other's reports against their own observations. Toward their peers, buyers attribute high willingness but lower competence, separating capability from intent. However, buyers do not make this separation for sellers, lowering all three dimensions together after a deceptive sale. The code can be found at: https://github.com/tapri-lab/trustforge.