Credo: Declarative Control of LLM Pipelines via Beliefs and Policies

A novel framework named Credo has been presented in an arXiv paper (2604.14401v2) to tackle the issues of opacity and fragility in agentic AI systems. It conceptualizes semantic states as beliefs and governs actions through declarative policies. Unlike traditional frameworks that depend on imperative control loops, short-lived memory, and logic embedded in prompts, Credo employs a database-backed semantic control plane for adaptive, auditable, and composable operations. The paper illustrates these ideas within a decision-control context, where beliefs and policies guide essential execution decisions like model selection and retrieval. This framework is tailored for environments necessitating long-term, stateful decision-making amidst constantly changing conditions, emphasizing the need for adaptability to new information and the revision of earlier conclusions.