Claude'u
değişik LLM'ler ile entegre etme imkanı bulunduğundan ve bu LLM'lerin bazılarının model parametrelerin toplam boyutları Claude'u meydana getiren kodların toplam dosya büyüklüklerinden defalarca büyük olduğundan (onlarca hatta yüzlerce GByte), neyin oranını incelediğimize makaleden bakalım:
11.1 Design Philosophy (sayfa 27)
The values and design principles introduced in Section 2 predict an architecture that invests in operational infrastructure rather than decision scaffolding. The implementation confirms this: the architecture documented in Sections 3 to 9 is overwhelmingly deterministic infrastructure (permission gates, tool routing, context management, recovery logic), with the LLM invoked as a stateless completion endpoint. An estimated 1.6%
of the codebase constitutes decision logic, the remaining 98.4% is the operational harness. This ratio is not accidental.
The design principles documented in Section 2.2 underpin this approach: the harness creates conditions under which the model can decide well, rather than constraining its choices. This design runs counter to the dominant pattern in agent engineering, where frameworks such as LangGraph route model outputs through explicit graph nodes with typed edges, and systems like Devin pair multi-step planners with heavy operational infrastructure. Claude Code instead gives the model maximum decision latitude within a rich operational harness. The engineering complexity exists not to constrain the model’s decisions but to enable them. This layered architecture, where the model reasons and the harness enforces, raises the question of whether agentic coding tools are converging toward operating-system-like abstractions in which the core loop serves as the kernel and everything else constitutes the OS.
Siz nasıl anlarsınız bilmem, makine öğrenmesinin kasik programlamadan temel farkı, modelin mümkün mertebe kuralları verilerden kendisinin çıkarmasıdır. Kodun sadece %1.6'sının decision logic geri kalan %98.4'ünün de operasyonel altyapıyla alakalı olması, yukarıdaki "Claude Code gives the model maximum decision latitude within a rich operational harness" cümlesiyle uyumlu. Yani dil modelinin (LLM'in) yönlendirmelerine kodla yazılmış kurallarla müdahaleyi minimize etmeye çalışmışlar anladığım kadarıyla. -Gürsu