Recent efforts to integrate large language models with causal discovery often rely on inferring graph structures or injecting outputs as priors, which risks conflating textual associations with genuine causal evidence. The authors argue that agents should instead assist the workflow by inspecting data, retrieving context, and clarifying assumptions without supplying edges, orientations, or causal conclusions. They propose a principle ensuring that causal claims remain grounded in data, explicit assumptions, formal algorithms, diagnostics, and expert decisions. To instantiate this approach, they introduce causal-learn+, an online platform coordinating preprocessing, method recommendation, and interpretation within the causal-learn ecosystem. A case study on Big Five personality data demonstrates an agent-assisted pipeline that avoids treating language model unreliability as causal evidence. The platform is available at causallearn.com.