Values
Constraints before instructions
Set the boundaries of a solution, then let the agent search inside them.
Direct the agent with goals, qualities and boundaries. Assume it will do the least sensible thing the boundaries allow, and close that option off. On most technical matters the agent knows more than you, so ask it for options and use what you add to focus its search without biasing it.
The trap
Engineers are trained to prescribe the route step by step, so you specify details that don’t matter and miss the one that does. The agent has no room left to find a better answer.
When instructions are right
- Agents know little about some enterprise infrastructure, so there you take a stronger hand.
- A decision needs more control, such as a domain model you are still testing.
Where it shows in the grid
“Hand over goals, not tasks, as big as you can assess” and “Define and enforce the qualities you care about, or don’t expect them.”