This paper proposes BRIDGE, a new algorithm for agentic reinforcement learning that jointly optimizes large language models (LLMs) and retrievers, addressing the information-credit gap in existing ARL methods.