Stateful Inference for Low-Latency Multi-Agent Tool Calling

In three linesStateful inference architecture for multi-agent tool calling with persistent KV cache across turns, reducing cost from O(n_t) to O(Δ_t). 2.1× speedup on 6-turn workflows, 4.2× on 35-turn median vs vLLM/SGLang.

Summary generated by Claude — human-verified