5 Rules for Agentic Engineering

Ready to take the step from vibe coding to actual AI Agentic SWE?

Here are 5 rules. Not suggestions.

1. Use custom agents.

More context is not better. No one in software design puts everything in one giant file. You modularize. Same principle — break your system into focused agents with clear responsibilities.

2. DRYP — Don’t Repeat Your Prompt.

Traditional engineering has DRY (Don’t Repeat Yourself): as soon as you write the same code twice, refactor into a helper function. Same here. If you’re writing the same prompt twice, have your skills-creator make an agent out of it. You’ll be surprised how well agents start stacking like LEGO blocks.

3. Use plan mode.

Separate planning from execution. If you do both at the same time, you’re burning context on reasoning that should already be done. An architect doesn’t draw blueprints while pouring concrete.

4. Never make a skill yourself.

Have your skills-creator do it. You built it for a reason — it’s better at this than you.

5. Never interact with an agent directly.

Have your agent orchestrator do it. You built it for a reason — it’s better at routing than you.

The common thread: treat your AI system like a real codebase. Modularize. Don’t repeat yourself. Separate concerns. Delegate to specialists.

That’s agentic engineering. Everything else is vibe coding.