In Mixture‑of‑Experts (MoE) language models, the router usually selects and weights experts solely based on a token’s hidden state, which limits the contextual information it can use. We introduce Attention‑Aware Routing (AAR), which enriches the router with temporal and spectral features extracted from a sliding window of attention weights. These features summarize the model’s contextual state and are disentangled from the hidden state.\ \ During training we keep the base transformer completely frozen and update only the routing parameters, isolating routing as the sole variable. On the OLMoE GSM8K benchmark, AAR improves over a routing‑only SFT baseline by +3.37 percentage points.\ \ Beyond raw performance, we demonstrate that routing and attention form a coupled circuit: a routing change at layer $l$ propagates through the residual stream and amplifies attention sinks at layer $l+1$, reshaping attention without any direct modification to the attention mechanism itself.\ \ AAR also reduces long‑form generation divergence: incorrect answers become shorter while correct answers retain their original length.\ \ Crucially, AAR is highly depth‑sensitive. Applying it indiscriminately across all layers harms factual retrieval, whereas introducing it deeper in the network preserves gains in mathematical reasoning. This reveals a retrieval‑reasoning tension across depth, and layer‑selective AAR serves as a controlled probe of the routing‑relevant information carried by attention at different layers.\ \ Review: AAR injects attention‑derived spatiotemporal cues into the routing process, achieving a synergistic coupling that boosts performance while offering a fresh lens on internal information flow.