Multi-hop reasoning is essential for knowledge graph completion but remains challenged by semantic incoherence in RL-based methods. Existing pairwise strategies often generate spurious paths by ignoring relation dependencies, while hierarchical strategies suffer from error accumulation and limited candidate spaces. To address these issues, we propose a context-aware hierarchical link prediction framework (CHLP) that incorporates global semantic information to guide multi-hop inference. CHLP features two coordinated agents: the Relation Policy Agent (RPA), which encodes relational trajectories using LSTM and dynamically fuses them with global relation context via a novel context fusion mechanism; and the Entity Policy Agent (EPA), which aggregates global-local entity context and introduces an embedding-based temporary entity space to enrich candidate entities. Both agents interact iteratively over multiple steps to construct semantically coherent reasoning paths. Extensive experiments on five benchmark knowledge graphs demonstrate that CHLP significantly outperforms state-of-the-art models in both prediction accuracy and path interpretability.