AgentReputation: A Decentralized Agentic AI Reputation Framework

Agentic AI marketplaces for software engineering tasks like debugging and security auditing are growing, but existing reputation mechanisms fail. Agents can strategically game evaluations, competence doesn't transfer across tasks, and verification effort varies wildly. Current approaches from federated learning or blockchain can't handle all three problems together. Enter AgentReputation, a three-layer framework that decouples task execution, reputation services, and tamper-proof persistence, allowing each layer to evolve independently.

The framework introduces explicit verification regimes tied to agent reputation metadata, plus context-conditioned reputation cards that prevent inflated scores from bleeding across domains. A decision-facing policy engine enables resource allocation, access control, and adaptive verification escalation based on risk and uncertainty. The paper, accepted at FSE 2026, also outlines future research: verification ontologies, quantifying verification strength, privacy-preserving evidence, cold-start bootstrapping, and defenses against adversarial manipulation.