agent-native-architecture

Agent-native architecture powers applications by treating agents as loop-driven orchestrators that operate with atomic primitives to achieve user outcomes. It shifts system design to be prompt-centric, defining features through prompts rather than embedding business logic in tools.

To build agent-powered apps, define atomic tools like read_file or store_item, write a system prompt mapping UI actions to these tools, and run an agent loop that reads prompts, calls tools, and signals completion using a dedicated tool like complete_task.

Context injection works by loading real-time app state into a shared workspace and injecting it directly into prompts. This ensures the agent acts on current data and maintains parity between the user interface and the agent's available context.

Ensuring parity means every UI action has a corresponding tool the agent can invoke. This enables flexible feature composition, allowing the agent to chain primitive actions through prompts to achieve complex, emergent outcomes.

You should avoid embedding business logic in agent tools when adopting a prompt-centric system design. Instead, keep tools atomic and define features as sections within the system prompt to support safe self-improvement and prompt-driven evolution.