A new technique, Agent‑Conditioned Routing (ACR), lets a large language model produce two or more different correct solution approaches to each of 126 math and physics questions. ACR runs four parallel plan‑making paths through a frozen mixture‑of‑experts model while keeping the original token router unchanged. All four paths share the same prompt and use deterministic decoding. On the benchmark, ACR produced 37.3 problems with at least two distinct methods on the GLM 5.3 backbone, compared with 18.7 when using the strongest temperature setting. With the GPT‑5.6 Sol backbone, the counts were 45.0 versus 27.0. A blind review of 30 problems confirmed that ACR yields more multi‑method solutions, but the overall correctness fell: Qwen ACR with a small mask had 57.5 % accuracy versus 98.3 % for temperature sampling. The extra memory cost is only about 160 KiB per path, adding roughly 3 % to peak VRAM use, and latency remains similar to batched temperature sampling. Thus, ACR can increase methodological diversity in generated solutions, but users must weigh the reduced precision and modest hardware overhead.