agent-native-architecture_skill

Overview

This skill teaches how to design agent-native architectures where agents are first-class citizens. It explains core principles, tooling patterns, and execution practices for building systems in which features are outcomes achieved by agents operating in a loop. Use it to shift product thinking from coded features to agent-driven outcomes.

How this skill works

The skill inspects your app design against five core principles: parity, granularity, composability, emergent capability, and improvement over time. It guides you to expose atomic primitives (read_file, write_file, list_files, bash, complete_task, etc.), map UI actions to agent capabilities, and structure a shared workspace and context files for accumulated knowledge. It also provides an architecture checklist and routing for focused design tasks.

When to use it

• Designing a new app where agents should perform user-visible actions autonomously
• Refactoring an existing product so features become prompt-driven outcomes
• Adding multi-step autonomous workflows that must operate reliably and resume
• Building MCP tools or deciding when to add domain-specific vs primitive tools
• Enabling safe self-modification and progressive prompt refinement

Best practices

• Ensure parity: every UI action must be achievable via agent tools or composed primitives
• Favor atomic primitives over bundled workflows so agents can make judgments
• Design CRUD-complete tools for every entity to avoid capability gaps
• Use a shared workspace and a context.md file for accumulated state and history
• Provide explicit completion signals and partial-completion checkpoints for resume
• Document available capabilities in the system prompt and refresh dynamic context

Example use cases

• A notes app where users ask an agent to summarize, tag, and file notes using read/write file primitives
• A downloads organizer where an agent analyzes file contents and moves items until the folder is organized
• A weekly-review feature added by a prompt that composes list_files, read_file and judgment to suggest priorities
• A task automation layer that discovers latent demand and surfaces new features from user requests
• A self-improving toolchain that refines prompts and context.md over time without shipping code