Jason Says
Today's double play — awesome-claude-skills giving you 1000+ production-ready Skills to grab, and OpenForgeRL letting you train harness-native agents end-to-end — signals that the Skills ecosystem is rapidly shifting from 'know how to use' to 'know how to train.' Don't sleep on this inflection point.
Composio launched awesome-claude-skills, a curated index of 1000+ production-ready Claude Skills and plugins spanning Claude.ai, Claude Code, Cursor, Gemini CLI, and more. Developers can plug in pre-built skills to give agents real-world actions without building from scratch — the most comprehensive community Skills directory to date.
Agents like Claude Code rely on complex inference harnesses that existing RL training stacks can't natively handle. OpenForgeRL is an open-source framework that lets developers train harness-based agents end-to-end in any environment — meaning you can now fine-tune Claude Code-style agents on your own infra without closed-source dependencies.
Block (Square's parent) open-sourced Buzz, a self-hostable workspace where humans and AI agents share the same rooms in real time. Built on a relay architecture users fully own, it's a notable early entry in 'sovereign AI workspace' — distinct from browser-based co-pilots, targeting real team collaboration scenarios.
LinkedIn co-founder Reid Hoffman and Zynga founder Mark Pincus launched AI lab Prentis, now in talks to raise $100M. Their bet: automating routine computer tasks will soon surpass coding as AI's biggest use case — a contrarian but plausible thesis from two seasoned internet veterans.
Day-2 real-world report on Claude Opus 5: outputs are notably more detailed and thorough, but the model is highly prescriptive — constantly offering unsolicited advice. Community adoption is faster than prior Opus versions; deemed worth the premium. Kimi K3 open weights dropping soon could multiply the impact further.
The former AI Search lead at Kimi (Moonshot AI) left to found Liangpei, applying AI search-and-matching to marriage matchmaking — covering profile collection, smart matching, and communication assistance. JOY Capital's Xu Xin invested $2M angel, citing the team's superior grasp of 'AI-powered people-finding.' A rare deep-vertical AI consumer play gaining early traction.
Training multi-turn agents is expensive because every update requires fresh environment rollouts and teacher model calls. ReOPD reuses pre-collected teacher trajectories as replayed prefixes, letting the student model only sample online at critical steps — significantly cutting training compute costs for developers building and fine-tuning agentic systems.
FinanceComplexQA introduces a 2000-question benchmark testing AI agents on real industrial-grade financial documents with complex layouts. It reveals significant performance gaps between agents on genuine financial analysis tasks — directly useful for developers building finance AI applications who need to evaluate which models actually hold up under real-world complexity.
Subscribe for daily AI updates + free playbook
📘 Subscribe Free