GitHub's trending repositories this week expose a fracture in how teams actually build AI agents. Rather than adopting framework-agnostic LLM libraries, developers are rapidly shipping opinionated, task-specific orchestration layers. Obra's Superpowers (557 stars overnight) packages agent composition as a "software development methodology," not an API wrapper. Matt Pocock's Skills framework (1,894 stars in 24 hours) organizes agent capabilities as composable unit tests developers can reason about locally—stored directly in `.agents` directories. Santifer's Career-ops demonstrates the pattern concretely: it chains job-scraping agents, evaluation agents that score listings against an A-F rubric, CV-tailoring agents, and application trackers, orchestrated within Claude Code and similar CLI environments. These aren't middleware layers or vendor SDKs. They're scaffolding for specific problems, built by engineers tired of waiting for generic solutions.
The surge reflects accumulated frustration with the current agent toolkit landscape. LangChain's agent abstraction, dominant 18 months ago, treats agents as retrieval-augmented chatbots—suitable for Q&A but inadequate for multi-step workflows with state, error handling, and human-in-the-loop checkpoints. Developers shipping production systems need primitives for task decomposition, rollback logic, and tool composition that map directly to their domain. Career-ops solves this by treating the job search as a directed workflow: scrape, score, personalize, submit, track. Superpowers frames agency as a methodology—agents own responsibilities, not just respond to prompts. Skills treats agent capabilities as testable, composable functions developers version-control alongside application code. None require a new model or training loop. All three treat the LLM as a component, not the architecture.
The velocity of adoption—nearly 2,650 combined GitHub stars in a single reporting window—signals market demand for what amounts to a new abstraction layer: above transformer APIs, below application logic. Three distinct design philosophies are competing: workflow orchestration (Career-ops), capability composition (Skills), and methodology framing (Superpowers). Critically, none emerged from an established framework vendor. They were authored by practitioners solving immediate problems. If this pattern holds, the winner for agent composition won't be the most theoretically elegant system, but the one that maps most directly to how teams already decompose work. The standardization battle isn't happening in research papers—it's happening on GitHub, in real codebases, and the ecosystem winner will likely be whoever controls the primitives developers instinctively reach for when building their next multi-agent system.