The Inference Report

The trending repos tell a story about what developers are actually building with AI right now, and it's almost entirely about agents and the infrastructure to run them at scale. The Claude Code ecosystem dominates, strix, agency-agents, career-ops, superpowers, and ECC are all frameworks for building specialized agents or optimizing their performance. What's notable is that these aren't experiments. They're tools designed to ship: career-ops includes a Go dashboard and batch processing; agency-agents describes each agent as having "proven deliverables"; superpowers frames itself as "a software development methodology that works." The token optimization angle (caveman's 65% reduction) and the focus on agent harnesses and skill frameworks suggest developers have moved past the "can we build agents" question and are now grinding on the practical problems: cost, reliability, and composability.

The discovery layer shows where the underlying plumbing is being built. Langflow and Xinference solve the infrastructure problem, one for visual workflow design, the other for swapping LLM backends with a single line of code. Unstract and Haystack's integration packages address data extraction and pipeline integration, problems that matter most when agents move from demos into ETL workflows. LlamaFactory's 72k stars reflect genuine traction in fine-tuning, which is becoming table stakes for teams that can't rely on closed APIs. What's absent is striking: no major new model releases in the trending set, no new training frameworks. The investment has shifted downstream, toward orchestration, optimization, and deployment. This is the phase where infrastructure maturity determines which teams can actually ship.

Jack Ridley