The Inference Report

The GitHub trends today reveal two distinct but overlapping movements: AI agents are moving from experimental frameworks into production tooling, and developers are building infrastructure to make those agents actually work at scale. The first wave, represented by repositories like Claude Code and agent-skills, treats agentic systems as a new class of developer tool, something that lives in your workflow and handles routine tasks through natural language. These aren't research projects or proof-of-concepts anymore. They're shipping with terminal integrations, codebase awareness, and git workflows baked in. The second wave is about the unglamorous layer beneath: how do you actually run multiple agents reliably, share context between them, keep sensitive data local, and audit what they do. That's where security-audit-skill, cua, and the MCP-context-forge show real engineering maturity. These solve problems that only appear once you've deployed agents into production, cross-OS fleets, verified findings, fault tolerance at scale.

What's striking is that the infrastructure conversation has shifted from "can agents do this" to "how do we do this safely and efficiently." ZenML positions itself as a unified platform from pipelines to agents. OpenMed runs clinical NLP entirely on-device to keep patient data local, solving a real compliance problem. Deja-vu solves something more subtle: it gives multiple coding agents access to a shared memory built from your own session history, meaning a fix discovered in one agent becomes available in another without requiring API calls or embeddings. These aren't flashy features. They're the details that separate a tool you can actually deploy from one that only works in a demo. The document management and market data tools in the mix, paperless-ngx and OpenStock, suggest a parallel pattern: developers are building open-source replacements for expensive SaaS platforms, and they're gaining traction because the underlying capability is now commoditized. What matters now is whether the tool integrates well with your existing systems and whether you own your data.

Jack Ridley