SmoothAgent: Efficient Long-Horizon LLM-Based Agent Serving with Lookahead Context Engineering

This document is a batch of recent arXiv CS papers (July 2, 2026) covering systems and ML infrastructure advances. Key contributions include: SmoothAgent — a lookahead context-engineering model and runtime that precomputes segment-decomposable context transformations to eliminate transformation overhead and reduce time-to-first-token (TTFT) up to 11.9×; MosaicKV — dynamic two-dimensional KV-cache compression for very long-context LLM serving delivering up to 16× attention speedup, ~3× memory reduction and only ~1.8% accuracy loss; ELDR — expert-locality-aware decode routing for PD-disaggregatd