The Inference Report

The GitHub ecosystem is consolidating around two distinct problems: how to build with Claude Code as a platform, and how to orchestrate multiple agents at scale. The Claude Code guides and templates dominate the trending set because they solve an immediate friction point, developers need working examples faster than documentation ships. luongnv89/claude-howto and the various "ultimate guide" and "learn" repositories succeed because they're artifacts, not abstractions. They show the thing working, then explain it. oh-my-claudecode and hermes-agent take this further by addressing the next layer of complexity: teams don't deploy single agents, they deploy systems that coordinate multiple agents with different capabilities. This shift from "how do I use Claude Code" to "how do I orchestrate Claude Code" mirrors the maturation cycle we saw with Docker and Kubernetes five years ago, first people learn the tool, then they learn to compose tools at scale.

Voice and robotics are gaining serious traction as applied domains. VibeVoice and airi represent the consumer-facing voice layer becoming viable enough that people build full applications around it. More significantly, openpilot's position in the discovery set shows robotics workloads moving from research labs into production systems managing real vehicles. The infrastructure layer supporting these workloads, skypilot for compute orchestration, PageIndex for RAG without vector databases, claude-mem for persistent session context, suggests developers are solving the plumbing problems that let agents run reliably in production. OpenBB and Twenty occupy the same space: financial data and CRM platforms that were once proprietary black boxes are now open source, and the differentiation has shifted from data access to agent-native APIs that let AI systems query and act on that data directly.

Jack Ridley