The Inference Report

The GitHub trending set reveals a decisive shift toward agent infrastructure and the tooling required to deploy it. What's moving is not individual models or frameworks but the operational layer: version managers like nvm sitting alongside specialized agent platforms, RAG systems designed for code understanding, and skill libraries that treat AI capabilities as composable, versioned artifacts. The pattern is unmistakable, developers are treating agents as production systems, not experiments.

Two categories dominate. First, agent orchestration and skill management. Anthropic's Skills repository, Addy Osmani's agent-skills, and Paperclip all solve the same fundamental problem: how to package, version, and deploy discrete capabilities so multiple agents can use them reliably. These aren't frameworks trying to abstract away the problem; they're registries and runtimes that acknowledge agents need persistent, reusable skill sets the way traditional software needs libraries. Orca and Paperclip go further, adding fleet management and desktop interfaces, treating agent deployment as an operational concern comparable to managing servers. Second, context and retrieval infrastructure. LightRAG, code-graph-rag, and Semantica tackle a harder problem: how to give agents accurate, traceable access to large codebases and knowledge without hallucination or context collapse. Code-graph-rag specifically solves monorepo navigation through knowledge graphs, a problem that didn't exist at scale until agents started writing code. These aren't replacements for existing RAG systems; they're specialized variants built for the constraints agents actually face.

What's notably absent from the trending set is framework churn. No new training paradigms, no new model architectures. What's present instead is the unglamorous infrastructure that makes agents reliable: version management, skill packaging, graph-based retrieval, and benchmarking systems like Harvey Labs that measure whether agents can actually handle domain-specific work. This suggests the market has moved past the question of whether agents work and toward the question of how to operate them. The discovery repos reinforce this: Distilabel focuses on synthetic data pipelines for training, AgileRL on hyperparameter optimization for RL, and EDSL on running agent-based experiments at scale. Developers are investing in the systems that make agents production-ready, not in the agents themselves.

Jack Ridley