ContReAct: A Feedback-Based Architecture for Continuous Agentic Operation

Abstract

We introduce ContReAct (Continuous Reason and Act), a reusable architectural pattern for extending the ReAct framework toward long-running, continuous operation of language model agents. Unlike traditional ReAct agents, which terminate upon task completion, ContReAct agents operate indefinitely through a self-feedback mechanism that enables sustained knowledge accumulation over extended periods. Our approach combines three mechanisms: (1) a reflection-plan-execute cycle with self-feedback that maintains temporal coherence without external objectives, (2) semantic deduplication for efficient memory management, and (3) history compression that preserves coherence across unbounded execution. The architecture readily supports diverse applications—from literature monitoring to knowledge base construction—by delegating adaptation to prompt engineering rather than architectural redesign. We implemented this approach and evaluated it on two case studies in which ContReAct agents autonomously build knowledge bases via web search on two subject topics, demonstrating accurate question answering and appropriate selection between memory retrieval and extending the knowledge base with new web searches.