The Inference Report

The trending set reveals two converging pressures reshaping how developers deploy AI. One wave treats agents as infrastructure: platforms like InsForge and goose abstract away the plumbing so you can bolt an AI coder onto your stack without reimplementing auth, storage, and compute from scratch. These repos solve a real friction point. The other wave optimizes the agent itself. Hmbown's DeepSeek-TUI, addyosmani's agent-skills, and the emergence of specialized tools like local-deep-research and PageIndex suggest developers are moving past "throw an LLM at the problem" toward building agents that search better, reason more efficiently, and fail more gracefully. The distinction matters: infrastructure repos lower the barrier to deployment; capability repos raise the ceiling on what agents can actually do.

What's notable is the absence of hype-driven clustering. You're not seeing dozens of "AI everything" wrappers. Instead you see domain specificity: TabPFN for tabular data, Whisper-Finetune for speech, docuseal for document workflows, OpenMontage for video. The discovery repos push further into this pattern. Kiln stacks multiple capabilities (evals, RAG, agents, fine-tuning) under one roof but doesn't pretend they're interchangeable. Unsloth's web UI for local model training and the emergence of multi-agent orchestration platforms like AgenticX and Kiln-AI suggest developers are moving toward composition over monoliths. They want to pick their LLM provider, their agent framework, their eval strategy, and their deployment target independently. That's the inverse of the framework-as-worldview trap. It's also where the real work is.

Jack Ridley