vLLM Semantic Router introduces a micro-agent runtime called the 'looper' that turns a single model API call into a bounded multi-model collaboration at the serving layer. Rather than requiring application-level agent graphs, the router itself orchestrates patterns like Confidence (sequential escalation), Ratings (parallel ensemble), ReMoM (mixture-of-model reasoning), Fusion (disagreement-based synthesis), and Workflows (planner/worker/verifier pipelines). The public surface remains a single OpenAI-compatible model name (vllm-sr/auto), while the router selects the appropriate recipe based on task difficulty, risk, and output contract requirements. Benchmark results show this approach matches or exceeds frontier single-model baselines on LiveCodeBench (92.6), GPQA-Diamond (96.0), and Humanity's Last Exam (50.0), beating Sakana Fugu Ultra and GPT-5.5 on several tasks. The core argument is that collaboration should be an open serving primitive, not a commercial endpoint or bespoke app logic.
