Why this matters

A normal chat answers the question you asked. An agent is given a goal and picks its own steps to get there. That is the power, and precisely the risk.

The 20-minute experiment

1. 1Think of a goal with several steps, like 'plan a two-day trip to Lisbon in October under 500 pounds'
2. 2Paste the prompt below with your goal, so it plans before it acts
3. 3Read its numbered plan and change one step before approving
4. 4Let it finish, then compare the result against the goal you actually asked for

Here is a goal. Before doing anything, show me your plan as numbered steps and wait for my approval.

• Goal given as an outcome, not a list of steps
• Numbered plan shown before any work started
• You changed at least one step
• You can name one thing it assumed that you never said