The Inference Report

The GitHub ecosystem is bifurcating along two clear lines: infrastructure for AI agents and tooling to manage the complexity they create. On one side, developers are building the plumbing agents need to work effectively. Repos like code-review-graph and wigolo solve a concrete problem: AI coding tools drown in context. Code-review-graph builds a persistent map of your codebase so language models read only what matters, reducing context overhead on large-repo workflows. Wigolo does similar work for web research, offering local-first search and crawl without API keys or cloud costs. These aren't viral because they're clever; they're gaining traction because they address a real friction point in agent workflows. On the other side, PostHog and opik represent the observability layer. PostHog bundles analytics, session replay, error tracking, and AI observability into one platform designed explicitly for products that use agents. Opik focuses narrower: tracing, evaluation, and monitoring for LLM applications and agentic systems. Both recognize that deploying agents at scale requires visibility into what they're doing and why they fail.

The other strong signal is standardization work. Apache Ossie tackles semantic metadata exchange across analytics, AI, and BI platforms, establishing vendor-neutral definitions so different tools can speak to each other without translation layers. This is unglamorous infrastructure, the kind that matters more than it trends. Meanwhile, the learning-to-build pattern persists: codecrafters and ai-engineering-from-scratch remain dominant because developers still want to understand what they're deploying, not just use it. AirLLM's ability to run 70B inference on a 4GB GPU matters for accessibility, but the real momentum is in the layers above raw inference: context management, observability, and semantic standards. The field is maturing past "can we run this model" toward "can we run this model reliably in production, and do we understand what it's doing."

Jack Ridley