Rohan Paul (@rohanpaul_ai)T2·78 pts
- Splitting search planning, code writing, and scoring into three separate owners avoids the high token cost of letting one coding agent steer the search.
- Hiding the scorer from the coding agent curbs reward-hacking and is the key design choice for efficient code-optimization loops.
- Solver quality rose 13.2% on the 680K-line C++ OR-Tools repo versus -0.1% for one-shot runs, showing iterative evolutionary search still drives the gains.