The creative reasoner before execution
"The old system could execute 'place a chair' but had no path at all for 'create a restaurant.'"
It was pure routing: input arrives, pick an agent, forward. Nothing decomposed intent before execution. The fix was diagnosing a missing capability, not a missing rule: a dedicated planning agent upstream whose only job is creative interpretation, so the tool agents stay dumb and fast.
// an architectural gap, not a prompt-tuning problem
Context before instructions
"Teach the agent who it serves and what breaks downstream before you tell it what to do."
Its system message is a curriculum, not a spec sheet: identity, then consequences, then reasoning patterns, then the instructions. Without that framing even a capable model reasons worse.
// a system prompt as a curriculum
Reference interpolation: "add trees" vs "move these trees"
"One paragraph killed a whole class of bugs where the model created new assets instead of touching the ones the user meant."
The fix taught reference resolution as a reasoning step before planning: interpret what "these" points to, then plan. Correct the thinking pattern, not the output.
// shipped against real model behavior