Today on GitHub trending, a pattern stands out: "skills" — reusable, composable agent capabilities — are emerging as a standard abstraction across the ecosystem. Cloudflare, Matt Pocock, and Google all shipped skills repositories to the trending chart on the same day, suggesting this layer of agent tooling is crystallizing quickly. Alongside that signal, NVIDIA's new graph reasoning library and continued Zed momentum round out a productive Saturday for open source.
cloudflare/security-audit-skill: Automated Multi-Phase Security Reviews
Cloudflare published a coding-agent skill that runs structured security audits against a codebase, producing verified, machine-readable findings — earning roughly 2,548 stars on its first trending day. What makes it notable is the design: rather than a one-shot scan, it uses a multi-phase approach — review first, then verify — with a schema that agents can consume programmatically rather than paste into a report. It's also a live demonstration of how Cloudflare's internal tooling is being built to interoperate with AI agents natively rather than bolted on after the fact.
github.com/cloudflare/security-audit-skill
mattpocock/skills: An Engineer's Personal Agent Skill Collection
TypeScript educator Matt Pocock published their personal skills directory, which gained 2,070 stars today. The repo is less a framework and more a curated reference — a window into how a working engineer structures reusable capabilities for their own coding agents. Its value is documentary as much as technical: it surfaces what kinds of tasks are worth encoding as skills versus handled inline, and the naming conventions signal what the community is settling on as the right level of abstraction.
NVIDIA/auto-ontology: Graph-Based Reasoning for LLM Retrieval
NVIDIA's auto-ontology is a Python framework that builds a graph reasoning layer on top of LLM and agent workflows, improving context retrieval and query generation — picking up 2,015 stars today. The core argument is that flat vector indexes lose structural relationships between concepts, while a graph layer preserves them so agents can navigate domain knowledge more precisely. For teams building retrieval-augmented systems where accuracy matters more than throughput, adding a semantic ontology layer rather than relying purely on embedding similarity is a compelling trade-off.
github.com/NVIDIA/auto-ontology
zed-industries/zed: Multiplayer Editor Keeps Building Momentum
Zed, the Rust-based code editor built for performance and real-time collaboration, added 306 stars today and continues a long run on the trending chart. The editors gaining traction in 2026 tend to share a few traits: built in systems languages, designed for the multi-agent workflow era, and treating collaboration as a first-class concern rather than a plugin. Zed fits all three, and for developers frustrated with Electron-based editors at scale, it remains the most credible alternative with active, opinionated development.