Reading up on DeepSeek V4 Pro

LibraryDesignBench evaluates agent‑generated libraries by measuring how much they reduce code for downstream agents, revealing that agents mimic human designs but improve when explicitly guided to optimize for agent usage.

The author describes building Pi Pocket via agentic vibe‑coding, noting rapid development but over‑engineered code, generic styling, and the need for manual tweaks.

Harness Handbook provides a behavior‑level manual that maps agent harness actions to verifiable code, enabling understanding, auditing, and safe modification.

OpenRouter's Fusion, which synthesizes outputs from multiple language models via a judge model, outperformed individual frontier models including GPT-5.5 and Claude Opus 4.8 on deep research tasks, with a budget panel achieving near-frontier performance at half the cost.