⚠️ Key Anti-Patterns (High-Yield Exam Content)
These appear as distractors in ~60% of questions. If you can spot them, you eliminate 2–3 options immediately.
| Anti-Pattern | Why It's Wrong | Correct Approach |
|---|---|---|
| Parse NL for loop termination | Brittle, non-deterministic | Check stop_reason value |
| Arbitrary iteration caps | Misses completion signal | Condition-based termination |
| Flat multi-agent architecture | No coordinator = no control | Hub-and-spoke pattern |
| Prompt-based enforcement | Probabilistic, bypassable | PostToolUse hooks |
| Self-review same session | Retains original reasoning | Separate session/subagent |
| Generic retry message | No context for correction | Append specific errors |
| Monolithic CLAUDE.md | Hard to maintain at scale | Hierarchical @import structure |
| All tools in one agent | Decision paralysis | 4–5 scoped tools max |
Study Tips
Focus on the WHY
The exam tests architectural reasoning, not memorization. Know why hub-and-spoke beats flat multi-agent. Know why hooks beat prompts for enforcement. Know why multi-pass review requires session isolation.
Anti-Patterns Are Traps
Exam distractors are often plausible anti-patterns. If an answer sounds like "just use a prompt for that" — it's probably wrong.
Build Real Projects
Hands-on experience with the Agent SDK, MCP servers, and Claude Code solidifies conceptual understanding faster than re-reading notes.
➡️ Next: 🔗 Resources & Sources.