Today's top repositories share a common thread: developers aren't just adding AI features to existing tools — they're rebuilding professional workflows from scratch with AI agents as first-class participants. Whether it's assembling a full marketing department from specialized agent personalities, editing video through conversation, or giving a design system APIs that work equally well for humans and AI assistants, the pattern is consistent. The tooling layer for AI-native work is forming fast.
232 Specialists, One Shell: Agency Agents
msitarzewski/agency-agents topped today's trending list with 2,114 new stars, offering a library of 232 personality-driven AI agent specialists organized across 16 professional divisions — marketing, engineering, security, sales, product, and more. Each agent ships with defined workflows, domain expertise, and measurable success criteria rather than generic prompt templates. The project bets that future AI-assisted work will look less like chatting with one general assistant and more like routing tasks to the right specialist in a fully-staffed virtual organization.
github.com/msitarzewski/agency-agents
Meta's Astryx: A Design System Built for Agents Too
Meta released Astryx as open source (708 stars), a design system that originated internally and grew to power over 13,000 applications before going public. What distinguishes it from established alternatives is an explicit architectural goal: the component APIs and documentation are structured so that AI assistants can build with them the same way a human developer would, rather than AI tools having to reverse-engineer conventions from existing code. With 150+ accessible components, no styling lock-in, and a CLI toolchain, it signals that "agent-ready" is becoming a serious design criterion rather than an afterthought.
Video-Use: Edit Video by Describing It
The browser-use team's video-use (693 stars) reimagines video editing around language rather than timelines: drop footage in a folder, then describe what you want in plain text. The technical core is that the agent works from a compact word-level transcript (~12KB) instead of processing raw video frames, which keeps token costs low while still enabling precise cut-point editing, color grading, subtitle generation, and animated overlays. It's a clear example of what AI-native tooling looks like when the interface is designed around how AI actually works rather than adapted from legacy GUI patterns.
github.com/browser-use/video-use
Vibe-Trading: A Multi-Agent Research Desk for Retail Investors
HKUDS/Vibe-Trading (694 stars) converts natural-language finance questions into backtested trading strategies, pulling from 18 data sources across US, Hong Kong, and A-share equities, crypto, futures, and forex. The platform deploys 29 pre-configured analyst teams — modeling the structure of an institutional research desk — and can export generated strategies directly to TradingView, MetaTrader, and similar platforms. A "Shadow Account" feature lets users import their own broker transaction history and compare their actual trading behavior against what a rule-based version of their strategy would have produced.