The Inference Report

The GitHub trending set reveals a decisive shift in how developers are building with AI: the focus has moved from model access to agent infrastructure. Where six months ago the conversation centered on which LLM API to call, today's most-starred repos solve the problem of what to do once you have that API. Browser-use and ponytail tackle agent reasoning and action, one by making web interfaces machine-readable, the other by teaching agents to avoid unnecessary work. Archify handles the output problem, generating verifiable diagrams from agent reasoning. The pattern is clear: developers are no longer asking "can I use Claude?" but rather "what can I build that Claude can actually do?"

The secondary pattern is ownership. Repositories like ai-job-search, claude-obsidian, and openhuman are explicitly forkable, self-hosted frameworks that let developers own their AI systems. Free-claude-code and the various plugin directories (claude-plugins-official, awesome-agent-skills) are less about discovery and more about reducing friction to deployment. Scientific-agent-skills is particularly notable because it's solving a vertical problem with validated skills, 175,000 scientists using a curated library suggests the market has moved past generic agents to domain-specific tooling. Vellum-assistant and the agent skills aggregators indicate that the real value now sits in orchestration layers and skill composition, not in the models themselves. The discovery repos like codebase-memory-mcp and defradb point to infrastructure problems that emerge once agents start operating at scale: how to make large codebases queryable and how to federate data across systems without centralization. These are the unsexy problems that determine whether an agent system works in production or becomes a prototype.

Jack Ridley