The Inference Report

The GitHub trending data reveals a decisive shift: developers are treating AI agents not as experimental features but as production infrastructure. The majority of high-traction repos fall into three overlapping categories. First, agent frameworks and orchestration layers, agency-agents, omnigent, herdr, that treat multiple LLM providers and specialized agent personas as composable components rather than monolithic services. Second, infrastructure for making agents practical: sandboxing (CubeSandbox), authentication (logto), token optimization (OmniRoute's compression saving 15-95% of tokens), and video editing via agents (browser-use/video-use). Third, tooling to extract and prepare data at scale, olmocr for linearizing PDFs into LLM-trainable formats, CVAT for vision annotation, the exercises dataset. What's notably absent from the trending set is any single dominant LLM provider's wrapper. Instead, repos prioritize abstraction layers that let developers swap models without rewriting code, which suggests the market has already decided that provider lock-in is a solved problem worth avoiding.

The discovery repos confirm this pattern while adding texture. SGLang and TanStack/ai both emphasize type safety and streaming primitives across multiple providers, they're solving the actual engineering problem of building reliable, testable AI applications rather than demo-ware. ARIS takes a narrower but revealing approach: autonomous ML research through markdown-only skills and cross-model review loops, no framework lock-in. Project Tapestry and Foundation-Models-Framework-Lab represent a secondary trend: developers building on constrained or regional models (sovereign models, Apple's on-device frameworks) rather than assuming global API access. The practical detail worth noting is that repos solving token efficiency, sandboxing, and annotation at scale are gaining serious traction alongside the agent frameworks themselves. This suggests the bottleneck has shifted from "can we build an agent" to "can we afford to run it, keep it safe, and feed it good data."

Jack Ridley