Mini agent loop lab

A small lab implementing plan → tool → observe → answer with hard permission gates and a max-step budget.

Agents feel magical until you implement stop conditions and approvals.

Register 1–2 safe tools (calculator, local search over a folder) → loop with step limit → require approval for anything external → log trajectory → score success.

Add a tool that is deliberately denied; confirm the agent cannot call it.

⚠️ Myth: More tools make the lab better.

✓ Reality: Two safe tools teach more than a junk drawer.

• Course 10 agents/tools/permissions; Course 21 API lab
• OWASP LLM Top 10: https://owasp.org/www-project-top-10-for-large-language-model-applications/ ↗