MORI, proposed by UC Berkeley, is a KV Cache Offloading solution for Agentic systems. The core insight is that "idleness is not a binary switch, but a relative continuous spectrum" — by sorting all active requests by idleness, the busiest stay on GPU while the idlest move to CPU, achieving 20-71% throughput improvement.

GREYHOUND is a USENIX ATC 2025 work that provides the first systematic characterization of fail-slow issues in production clusters with 10,000+ GPUs, and proposes a complete solution with non-invasive detection and ski-rental-inspired multi-level mitigation. Detection accuracy exceeds 99%, end-to-end throughput improves by 1.58x.

Found a tool called lark-channel-bridge that connects Feishu directly to your local Claude Code — send a message from your phone, and Claude runs code, edits files, deploys your site on your computer. Tried it, works well, documenting the process and pitfalls.

From OpenAI's million-line code experiment to deep discussions on Zhihu, Harness Engineering has become the hottest topic in the Agent era. But rather than talking about "concepts," I want to share the pitfalls I hit while using Claude Code on real projects—and how those pitfalls gradually forced me to put reins on my own Agent.

A deep dive into the core technical innovations of DeepSeek-V4-Pro and V4-Flash, covering hybrid attention with HCA, contextual sparse attention CSA, Contextual Parallelism training strategy, inference optimization, and other key engineering practices.

Anthropic released Claude Opus 4.8 alongside Dynamic Workflows—a feature that lets the model write JS scripts on the fly to orchestrate dozens to hundreds of sub-agents for super-large tasks. This post takes a deep look at its architecture, execution model, real-world cases, and when to reach for it.