Agent-Native Architecture

Most apps treat agents as a chatbot bolted onto existing code. Agent-native apps invert that: features are outcomes you describe, achieved by an agent with tools, operating in a loop.

This skill teaches you how to design applications this way — from principles to implementation.

The 5 core principles:

• Parity: Whatever users can do in the UI, agents must be able to achieve through tools
• Granularity: Prefer atomic primitives; features are outcomes, not functions
• Transparency: Surface agent state so users can trust and verify
• Controllability: Users must be able to pause, redirect, and correct at any step
• Reversibility: Design for undo — agents make mistakes

What you get:

• The parity test: audit whether your app is truly agent-native
• Capability map templates (UI action → agent tool)
• MCP tool design patterns that work at scale
• Self-modifying system patterns (agents that improve themselves)
• Anti-patterns to avoid: god tools, silent failures, state leakage

Applies to Claude Code plugins, MCP servers, OpenClaw agents, Paperclip workflows, and any agentic system.

Use before designing your next feature. The question to ask every time: can an agent achieve this outcome?