Today's GitHub trending is a clean signal about where the developer community's attention has landed in mid-2026: not new foundation models, but the scaffolding around them. Self-improving agents, specialist role frameworks, and reusable skill libraries are crowding out the experiments of a year ago. The infrastructure layer is becoming a real thing you install, version, and maintain — not just prompts you paste into a README.
Prime Agent: A Coding Agent That Improves Itself
PrimeIntellect's prime-agent is the biggest mover today with over 2,300 new stars. Written in TypeScript, it targets long-running autonomous coding workflows and introduces a self-improvement loop built on reinforcement learning from model feedback. Rather than completing a task and stopping, the agent evaluates its own output, flags what went wrong, and updates its behavior for the next run. For teams trying to run coding agents at scale without constant human correction, this feedback architecture is a meaningful step beyond simple task execution.
github.com/PrimeIntellect-ai/prime-agent
Agency Agents: Specialized Roles for AI Teams
msitarzewski/agency-agents picked up nearly 900 stars today by building out the multi-agent model as a full simulated agency — complete with distinct specialists like frontend developers and community analysts, each with their own defined personality and scope. The project makes the case that narrow specialization beats generalism even in language-model-based systems, and it provides a working scaffold for developers who want to spin up a team of role-specific agents rather than a single do-everything assistant.
github.com/msitarzewski/agency-agents
Addy Osmani's Agent Skills: Engineering Discipline, Packaged
Google's Addy Osmani published agent-skills, a JavaScript library of production-grade engineering behaviors designed to be composed into AI coding agents. With nearly 700 new stars on its first trending day, the project fills a gap that most coding agents still have: the difference between writing syntactically valid code and following real engineering discipline. The skills cover error handling, testing habits, documentation standards, and similar practices that senior developers internalize but that agents need to be explicitly taught.
github.com/addyosmani/agent-skills
Google's Official Agent Skills Library
Running alongside Osmani's independent effort, Google published a separate skills Python library — a curated set of agent capabilities for its own products and developer ecosystem, which gathered over 500 stars today. The parallel release from two different parts of Google on the same day suggests the company is treating reusable agent skills as a first-class architectural concept: something you package, version, and distribute through normal dependency channels rather than embed ad hoc in system prompts.
witr: Find Out Why a Process Is Running
Not everything trending is agentic. pranshuparmar/witr is a Go-based CLI and TUI that answers the question every developer eventually faces: why is this process actually running? Given a process, port, container, or open file, it traces the dependency chain back to the original trigger. It earned over 200 stars today — a modest number by today's standards, but a sign that sharp, focused Unix-style tools still find an audience fast when they solve a problem that existing tools handle poorly.