The Inference Report

The GitHub trending set reveals a field consolidating around agent infrastructure and the tooling required to make agents reliable. Most of the high-velocity repos solve a specific operational problem: how to build, test, and deploy AI systems that do things rather than just generate text. Agency-agents and ai-hedge-fund treat agents as composable units with defined roles and outputs. Promptfoo addresses the upstream problem, testing those agents before they run loose, with the kind of declarative config approach that integrates into CI/CD pipelines. Superpowers frames agent development as methodology, not just library. These repos share a pragmatic constraint: they assume you're shipping something that needs to work, not experimenting with what's theoretically possible.

The discovery set splits into two movements. One handles the data layer: Chroma provides vector retrieval as a standalone service rather than bolted onto an LLM, LightRAG (published at EMNLP 2025) optimizes retrieval-augmented generation for speed and simplicity, and Pixeltable treats multimodal data as a first-class infrastructure problem with incremental computation. Label-studio remains the standard for the unglamorous work of preparing training data at scale. The other movement addresses control and safety. Semantic-router makes routing decisions without invoking expensive models. Moltis packages agent runtime as a single auditable binary with sandboxing built in. DataDesigner generates synthetic data programmatically, solving the bootstrapping problem when real data is scarce or sensitive. What's notably absent from the trending list: repos that wrap existing closed-source APIs. The investment is flowing toward systems you can own, test, and reason about end-to-end.

Jack Ridley