The Inference Report

The GitHub ecosystem is consolidating around two distinct problems: making AI agents practical, and making them controllable. The agent-focused repos dominate trending, but they split into infrastructure versus skills. On the infrastructure side, repos like trycua/cua and activepieces/activepieces solve the actual hard problem: how do you let an AI agent interact with a desktop or trigger workflows without it going rogue or hallucinating its way into production? These aren't frameworks asking you to rewrite your stack. They're sandboxes, benchmarks, and orchestration layers that treat agent control as a first-class concern. The skills layer, K-Dense-AI/scientific-agent-skills, mattpocock/skills, danielmiessler/Personal_AI_Infrastructure, packages domain knowledge as reusable components. This reflects where real development effort is going: not building agents from scratch, but teaching them to do specific work reliably. That's a maturation signal.

Separately, there's a strong current in local, on-device computation. supertone-inc/supertonic runs multilingual text-to-speech natively via ONNX. mlx-embeddings handles vision and language embeddings on Mac hardware without cloud dependencies. These aren't about performance theater. They solve a real constraint: latency, privacy, and cost for production systems. Meanwhile, rasbt/LLMs-from-scratch and github/spec-kit sit at opposite ends of the spectrum, one teaches you how LLMs actually work from first principles, the other automates spec-driven development workflows. Both are getting serious traction, suggesting developers want either deep understanding or high-level automation, not much in between. The infrastructure plays are winning because agents are only useful if they don't become liability machines.

Jack Ridley