SourcesDualPath DeepSeek Solves Agentic Inference BottleneckarXiv·high signalXBlueskyLinkedInCopy linkDeepSeek-AI presents DualPath addressing storage bandwidth bottleneck in multi-turn agent inference. 1.87x offline throughput 1.96x online serving improvement via RDMA dual-path loading.SourceSource pagearXiv↳ Follow the threadStack layer / ContrastReplace black-box top-K retrieval with interpretable agentic document navigation: 58.8% vs 15.7% on financial documentsarXiv 2608.06305Stack layer / ContrastRole Specialization Model Case Study Coordinates Antigravity, Gemini CLI and Local Qwen Code on One Python ApparXiv 2608.12311Policy dependency / Stack layer'Rethinking Agent Security as a Networking Problem': stop asking the LLM to enforce its own policyarXivStack layerLoad Hijack Poisons Only MoE Router Weights — 92.3–95.6% of Triggered Tokens Land on One GPU, Cutting Throughput to 0.86xarXiv 2608.10614Stack layer / Threat patternPattern: Opaque Client-Held State Is the Agent Stack's Newest Leak SurfacearXivStack layer / Contrast"Not Worth Another Token": Pruning Deep-Research Agent Context Before Retrieval Cuts Token Usage Up to 73% With Little Quality LossarXiv (via HuggingFace Daily Papers)Stack layer / ContrastContinuity Kernel: a transactional commit protocol for long-lived agent state, model-checked over 2.8M statesarXivStack layer / Update threadADEPT Unifies Five Families of Deep-Learning Test Adequacy Metrics Under One YAML-Configured WorkflowarXiv 2608.12144