Introduction
In an AI-native org, agents write most of the code - yuva/namzu and external coding agents alike. Left to itself, that is a liability. A large language model generates by statistical prediction: the same problem, asked twice, yields two similar-but-not- identical solutions. Most AI review tooling only sees the diff, so a freshly generated helper that already exists elsewhere is invisible to it. Unchecked, more AI means more divergence - a fourth Button, a near-duplicate date formatter, a service that reinvents a primitive three repos already have.
An AI-native org that is agent-first and full-scratch over one canonical model
- Cogitave's own is Cogitave Core - cannot treat divergence as a style nit; it undercuts the very thing that makes an estate harmonic. The counter-force is deliberate: retrieval before generation. Cogitave codifies that operating model in its AI-Native Development standard (decision: ADR-0013).
IMPORTANT
Before authoring any UI or code, every agent and every human MUST query the registry and the design system for an existing fit. Authoring something new is the exception, requires a recorded "no fit" justification, and is contributed back to the shared library - never kept as a local fork.
The same discipline applies one level up, to whole engines and tools, not just components: ADR-0003 governs when "build it ourselves" is actually earned versus when it is blind reinvention of a problem the industry already solved well.
What you will get from this module
- Why reuse-first, and what enforces it - the discover-before-generate hard rule, the tokens-to-primitives-to-golden-components single source of truth, the MCP-native registry, and the anti-duplication gates that keep a needless re-author out of
main. - Build from scratch as the exception - ADR-0003's port-best-in-class, rebuild-native stance: references are studied and ported, not reinvented blindly, and a from-scratch rebuild has to be materially earned.
By the end, you will be able to state the hard rule, name what enforces it, and justify a from-scratch decision the way an AI-native estate's own ADRs require - as a reasoned exception, never a default.