Today's GitHub trending page reads like a design specification for the next generation of agentic tooling. Four high-momentum repos share the same core premise: raw LLM capability isn't the bottleneck anymore — the gap is reusable skills, persistent memory, and principled workflows that let agent teams do real work over real timescales.
TencentCloud/TencentDB-Agent-Memory: Persistent Memory Across an Entire Agent Team
TencentCloud's new open-source project addresses a coordination problem that gets harder as agent systems grow: when multiple agents are collaborating, memory that lives inside a single session isn't shared, versioned, or reusable. TencentDB-Agent-Memory provides a centralized memory hub where agents can store, retrieve, and build on each other's findings — more like a team wiki than a chat history. Built in TypeScript and racking up over 1,000 new stars today, it's attracting attention from developers building multi-agent pipelines where context persistence is a genuine operational requirement.
github.com/TencentCloud/TencentDB-Agent-Memory
addyosmani/agent-skills: Engineering Best Practices as Agent Instructions
Addy Osmani's agent-skills repository packages decades of software engineering wisdom into a format AI coding agents can actually use. Rather than hoping a model has internalized good practices from training data, this project externalizes them as discrete, composable skill files — covering code review, performance profiling, debugging approaches, and architectural thinking. The repository has grown to over 83,000 stars, a sign that the gap between "what AI agents can do" and "what production engineering actually requires" is a pain point teams are actively trying to close.
github.com/addyosmani/agent-skills
mattpocock/skills: Skills Written for Engineers, Not for Demos
Matt Pocock brings a TypeScript-ecosystem perspective to the agent skills space, with a collection of shell-based skills targeting the day-to-day decisions real engineers face rather than synthetic benchmarks. The framing is explicit: these are skills for engineers who want their agent to behave the way a competent senior developer would, not a demo-optimized assistant. The project gained nearly 2,000 stars in a single day, suggesting strong resonance with developers frustrated by the gap between agentic marketing and agentic reality.
obra/superpowers: A Full Development Methodology for Agentic Work
Jesse Vincent's superpowers project takes the widest scope of the four, framing itself not just as a skills library but as a software development methodology built around agentic workflows. It defines how agents should approach tasks, escalate decisions, manage state across sessions, and hand off work — more like a team process document than a plugin. At 268,000 total stars and nearly 900 new today, it's striking a nerve with developers who want principled structure around agent-assisted development rather than ad hoc prompting.