Credo: Declarative Control of LLM Pipelines via Beliefs and Policies

arXiv cs.AI / 4/17/2026

📰 NewsDeveloper Stack & InfrastructureIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that agentic, long-lived LLM systems need correctness not just from individual outputs but also from how they adapt as new evidence arrives and prior conclusions are revised.
It criticizes existing agent frameworks for using imperative control loops, ephemeral memory, and prompt-embedded logic, which can make behavior hard to understand, brittle, and difficult to verify.
Credo is proposed as a system that represents an agent’s semantic state as explicit beliefs and controls actions via declarative policies defined over those beliefs.
The approach uses a database-backed semantic control plane to enable adaptive, auditable, and composable execution, allowing behavior to change without modifying the underlying pipeline code.
A decision-control case study shows how beliefs and policies can guide key choices such as model selection, retrieval, and corrective re-execution.

Abstract

Agentic AI systems are becoming commonplace in domains that require long-lived, stateful decision-making in continuously evolving conditions. As such, correctness depends not only on the output of individual model calls, but also on how to best adapt when incorporating new evidence or revising prior conclusions. However, existing frameworks rely on imperative control loops, ephemeral memory, and prompt-embedded logic, making agent behavior opaque, brittle, and difficult to verify. This paper introduces Credo, which represents semantic state as beliefs and regulates behavior using declarative policies defined over these beliefs. This design supports adaptive, auditable, and composable execution through a database-backed semantic control plane. We showcase these concepts in a decision-control scenario, where beliefs and policies declaratively guide critical execution choices (e.g., model selection, retrieval, corrective re-execution), enabling dynamic behavior without requiring any changes to the underlying pipeline code.