In agent-native architecture, features are prompts — not code

The old model treats features as functions you coded. The agent-native model gives an agent atomic tools and describes outcomes — features become prompts, and the agent loops until the outcome is achieved. This is the core of ai-native-product-architecture.

The key principles that make this work include parity (agent can do anything UI can), granularity (tiny primitive tools), composability (new prompts = new features), emergent capability (agent does things you never designed), and self-improvement. (Note: this framework synthesizes themes from Shipper’s writing; the “five pillars” label is editorial.) The composability principle means you build for obsolescence naturally — new capabilities ship without code changes.

The ultimate test: describe an outcome within your application’s domain that you didn’t build a specific feature for. If the agent figures it out, you’re agent-native. This connects to why Declarative beats imperative when working with agents — you define success criteria, not steps. Taken to its extreme, this becomes Malleable software — a tiny core that writes its own plugins — replaces fixed-feature applications — the application itself generates new prompts to match user behavior, extending its own capabilities without developer intervention.