The Inference Report

The trending cohort reveals a decisive shift: developers are building infrastructure for autonomous AI agents, not just using large language models as chatbots. Cloudflare's security-audit-skill and the Claude-Red offensive security library sit atop the list precisely because they treat LLMs as executable agents that can be equipped with domain-specific skills and then verified against real outputs. The pattern repeats across NVIDIA's OpenShell, which provides a runtime for autonomous agents, and Anthropics' knowledge-work-plugins, which extends Claude's capabilities into structured problem domains. These aren't wrappers around existing models. They're frameworks that let you compose specialized behaviors, measure their correctness, and run them without human intervention at each step. Node.js and PHP remain in heavy rotation because production systems still run on them, but the gravitational center has moved to the agent layer.

The discovery repos show where the practical problems live: vector databases like Qdrant and Lance are gaining traction because multimodal AI systems need fast approximate search over massive datasets, not just inference speed. Picollm and tt-metal address the opposite constraint, pushing inference onto devices where bandwidth and power matter more than throughput. Botpress and beta9 solve the deployment problem, letting teams package agents and run them without managing Kubernetes clusters. The smaller repos like DiffusionART and the multimodal reasoning handbook suggest that reinforcement learning is moving from theory into practice, though adoption is still narrow. The real signal is not in any single repo but in the infrastructure gap they fill: developers can now train agents, equip them with skills, verify their outputs, and ship them to production. Three years ago, that pipeline didn't exist.

Jack Ridley