🌐 日本語

Agents Design Principles

IMPORTANT

→ Why: Context Rot mitigation (execution in independent context window) → Why: Sycophancy mitigation (eliminate compliance bias via Cross-model QA) → Why: Knowledge Boundary mitigation (narrow knowledge domains to reduce boundary-crossing probability)

What Are Agents?

Agents are sub-agents operating in independent context windows. They do not consume the main context.

Attribute	Value
Invocation Timing	Called via Agent() / Task() tools
Context Consumption	Independent context window, separate from main
Appearance to LLM	Separate LLM instance with its own system prompt

Biggest Difference from Skills

Main Claude (context window)
  ├─ Skill loaded → executed in same context (consumes context)
  │               → equivalent to import / require
  │
  └─ Agent launched
       └─ separate context window (independent)
          → only result returned to main (distilled summary)
          → equivalent to delegation to separate process

Addressing Structural Problems

Context Rot Mitigation

Agents run in fresh, independent context, unaffected by main Context Rot. Even mid-session with long history, delegating to an Agent yields high-quality output.

Sycophancy Mitigation (Cross-Model QA)

Review with different models or fresh context enables verification without shared compliance bias.

Knowledge Boundary Mitigation

Delegate tasks requiring specific domain knowledge to specialist agents, narrowing knowledge domains and reducing boundary-crossing probability.

Agent Definition Example

<!-- .claude/agents/code-reviewer.md -->
---
name: code-reviewer
description: Specialist agent for code review of PRs
model: sonnet
allowed-tools:
  - Read
  - Grep
  - Glob
  - Bash(git diff *)
---

You are an Angular/NgRx code review specialist.

## Review Perspectives
1. NgRx pattern compliance
2. RxJS memory leaks (missing takeUntilDestroy)
3. Alignment with OnPush change detection
4. Test coverage

Important Constraints

• Sub-agents cannot invoke other sub-agents
• Indirect invocation via bash commands is also prohibited
• Always use the Agent(subagent_type="agent-name", ...) tool

---

Previous: Skills Design Principles

Next: Criteria for Skill vs Agent