The agent loop is read, decide, act. A plain chatbot cannot act, but it can decide out loud and then work through its own plan. That is most of the value of an agent for many tasks, and it costs one turn instead of twelve. It also shows you, before you buy anything, which steps would have needed a real tool.
| Prompt | What comes back | What you learn |
|---|---|---|
| "Plan a week of dinners for two on $60" | A list of meals, probably over budget, arithmetic unchecked | Nothing about how it got there |
| "First write PLAN as numbered steps. Then EXECUTE each step, showing your working. Mark any step where you would need a tool you do not have. End with RESULT." | A visible plan, a costed list, a total, and a note that it cannot check today's prices | Which steps are reasoning (free) and which are actions (would cost a tool call) |
Try it
Plan first, then execute
Now yours
The template below is the same pattern with your task in it. The slot for tools you lack is the important one: a model told it has no calendar will write "you should check the calendar" instead of inventing a free slot on Tuesday.
Your turn
Your own plan-then-execute prompt
Act like an agent with no tools. First write PLAN: numbered steps to complete the task. Then EXECUTE each step in order using only the information I give you, and under any step that would need a tool ([TOOLS YOU LACK]), say so and use my information instead. Finish with RESULT. TASK: [TASK] INFORMATION: [INFORMATION]
Hint: Naming the tools it lacks is what makes it flag actions instead of pretending it did them.
Knowledge check
Your EXECUTE section says "Step 2: I looked up current supermarket prices and found chicken at $8.50/kg." You gave it a price list with chicken at $9. What happened?