The Inference Report

The GitHub trending set splits cleanly between two currents: infrastructure for AI agents and practical tools for converting, storing, and moving data. The agent side is where most of the energy sits. Claude Code dominates by a wide margin, but that's partly because Anthropic ships it as a standalone tool people install. More revealing are the supporting structures: Harness designs domain-specific agent teams, CORAL builds infrastructure for multi-agent self-evolution, and Pi Subagents handles async delegation with session sharing. These repos solve a real coordination problem. Once you deploy multiple agents, you need ways to compose them, track their decisions, and keep humans in the loop. Aegis addresses the control layer directly, enforcing runtime policies with cryptographic audit trails and kill switches. None of these existed two years ago because the problem didn't exist at scale.

The other current is less flashy but more widely useful: MarkItDown converts documents to Markdown, Scrapling handles web scraping at any scale, and Supermemory builds a memory layer for AI systems. These are glue tools. They solve the friction of getting data into and out of AI pipelines without forcing you into a specific framework. Apache Hamilton and ModelScope operate in the same space, Hamilton gives you testable, lineaged dataflows that run anywhere Python does, while ModelScope treats models as a service you can call. The discovery repos cluster around safety and finance. Aegis and CORAL both assume agents will make decisions that need auditing or reversal. Awesome AI in Finance is a curation, not a tool, but its existence signals that people are building financial applications on top of LLMs and need to know what's already been tried. Openpilot remains the most interesting robotics repo on GitHub because it ships actual code running on actual cars, not research code or simulation. The pattern across all this: developers are moving past "what if we use an LLM" and into "how do we actually operate these systems."

Jack Ridley