Reference implementations and minimal working examples are becoming a dominant pattern for sharing knowledge in agentic workflows. These might be called pseudo-frameworks.

Before agents, frameworks existed partly because the alternative — handing someone an example and saying “build yours like this” — imposed too much cost. The edges between content and framework are muddy, the surface area large. Unpicking someone else’s plumbing required exactly the kind of long uninterrupted cognitive work that humans find expensive and often avoid. A framework abstracted that away at the cost of lock-in.

Now the disentangling is cheap. You point an agent at someone else’s plumbing, nudge it along, and get something that maybe works. The framework’s main value proposition — saving humans from reading work — is largely gone. What’s left is the lock-in.

Other names

  • Reference implementations
  • Starter repos / starter kits / starter templates
  • Blueprints (NVIDIA)
  • Copy-paste templates / copy-this-repo patterns
  • Scaffolds / minimal working examples
  • AGENTS.md style files

Sources and discussions

  • Microsoft Fabric — Operationalizing Agentic Applications — “production-minded reference implementation” (Agentic Banking App) as the pattern; not a framework, just a concrete artifact to study and fork.
  • NVIDIA Build platform — Retail Agentic Commerce Blueprint — “Blueprints” positioned as reference implementations to adapt, not frameworks to hook into.
  • Claude Code / Cursor communities (GitHub, LinkedIn, Instagram) — heavy sharing of “copy this repo” multi-agent team repos (e.g., msitarzewski/agency-agents). Explicit contrast with heavy frameworks.
  • DeepLearning.AI / GitHub threads on Agentic RAG, LangGraph, Claude Agent SDK — “Not another framework — just a working example you can fork.”