DeepTutor attempts something most AI tutors skip: combining retrieval-anchored explanations with stateful memory of what you've mastered. The HKUDS team supports LlamaIndex, GraphRAG, Obsidian, and five other backends while tracking learner progress across sessions—a technically ambitious architecture that's still working through orchestration costs, interface instability, and performance bottlenecks.

Cursor hits Claude's rate limit. You switch to OpenAI manually. Then Gemini. Then back. OmniRoute turns this dance into a single OpenAI-compatible endpoint at localhost:20128 that routes across 20+ providers with quota-aware fallback. Built from 9router and CLIProxyAPI DNA, battle-tested with 83K stars and real deployment scars.

AI coding agents excel at functional UI code but consistently make poor design choices—ease-in animations for entrances, solid borders instead of subtle shadows. Emil Kowalski's skills repository addresses this gap with design guidelines that teach agents aesthetic judgment, gaining 8,000 GitHub stars in one week despite early growing pains with factual errors and file length limits.

Cloudflare open-sourced the multi-agent security workflow they use internally, complete with 13,841 production findings and transparent benchmark data showing 15× higher costs than single-agent reviews. The release includes their six-phase architecture designed to reduce false positives through adversarial validation and independent verification.

Vectorize.io's Hindsight tackles persistent memory for AI agents with entity tracking, temporal reasoning, and multi-strategy retrieval—outperforming RAG baselines by 58% in benchmarks. Early adopters report real wins in coding agents and incident triage, though database contention and concurrency remain active engineering challenges.

AI agents generate scientific code fast, but scientifically defensible code requires validation protocols, authoritative identifiers, and domain conventions. K-Dense externalizes 177 skills spanning biology, chemistry, medicine, and physics into modular, reusable components that bridge the gap between runnable and publishable.