agent-native（Agent 本机）

Agent-Native

Agent-native applications treat agents as first-class users. Whatever a human can do through the UI, an agent can achieve through tools. Features are outcomes described in prompts, achieved by an agent with atomic tools, operating in a loop until done.

Reference Files

Core Principles

• Parity — Agent can achieve anything users can do through the UI. Build a capability map; close every gap.
• Granularity — Tools are atomic primitives. Judgment and decision logic live in prompts, not tool implementations. To change behavior, edit prose, not code. # Anti-pattern: workflow-shaped tool analyze_and_organize(folder) # bundles judgment into code # Agent-native: atomic primitives list_files(folder) → read_file(path) → move_file(src, dst) → write_file(path, content) # The agent decides what to move and where — judgment stays in the prompt
• Composability — New features = new prompts. With atomic tools and parity, you describe an outcome and the agent loops until it's achieved.
• Emergent capability — The agent handles requests you didn't design for. Observe what users ask; formalize patterns that emerge.
• Improvement over time — State persists via context files. Prompts can be updated for all users without shipping code.

Design Workflow

1. Capability audit — Map every UI action to an agent equivalent. Close any gap. Read core-principles.md.
2. Tool surface design — Define atomic primitives for every entity (CRUD). Add domain tools only for vocabulary, guardrails, or efficiency. Read tool-design.md.
3. Context and state planning — Design the context.md file, entity directory structure, and system prompt injection. Read files-and-context.md.
4. Agent-UI feedback design — Define completion signals, event types, shared workspace, and approval gates. Read agent-ui-communication.md.
5. Validate — Run the checklist below. Describe an unbuilt outcome and test whether the agent can figure it out.

Design Checklist

Copy and track during design:

Agent-native design progress:
- [ ] Capability map: every UI action has an agent equivalent
- [ ] Tools are atomic primitives (judgment in prompts, not tools)
- [ ] Every entity has full CRUD tool coverage
- [ ] System prompt injects available resources and capabilities
- [ ] Agents and users share the same data space
- [ ] Agent actions reflect immediately in UI
- [ ] Completion is signaled explicitly (no heuristic detection)
- [ ] External APIs use dynamic capability discovery where possible
- [ ] Approval model matches stakes and reversibility
- [ ] Ultimate test: describe an unbuilt outcome — can the agent figure it out?

Validation Loop

Before shipping, verify:

1. Capability coverage — Pick 5 random UI actions; confirm the agent can accomplish each without touching code.
2. Tool atomicity — Review every tool; if it contains if/else decision logic, split it.
3. Context completeness — Clear the agent's context and start a session; does it still know what exists?
4. End-to-end scenario — Give the agent an outcome you never explicitly built. Does it compose tools to get there?

Anti-Patterns

• Agent as router — Agent routes to your code instead of acting with judgment; you've built a dispatcher, not an agent.
• Workflow-shaped tools — analyze_and_organize bundles decision logic into a tool; break into read_file, move_file, write_file.
• Orphan UI actions — User can do something the agent can't; breaks parity.
• Context starvation — Agent doesn't know what exists; inject resources and capabilities into every system prompt.
• Gates without reason — Domain tool is the only path when primitives should also work; default to open.
• Heuristic completion detection — Detecting done via consecutive idle iterations; require an explicit completion tool call.
• Static API mapping — 50 tools for 50 endpoints; use discover + access instead.

Related Skills

• define-architecture — repo structure and module boundaries before going agent-native
• agents-md — audit CLAUDE.md / AGENTS.md for agent instruction quality