Gru & Minion
A frontier model, Gru, owns the task end to end and chooses when to delegate bounded work to a cheap model, Minion. Gru re-runs the checks itself instead of trusting the agent’s report. I benchmarked it on the real SWE-bench and GAIA harnesses with model pairs from three vendors, and all three matched or beat their solo baseline on SWE-bench.
- 92.3%
- of tokens moved to the cheap model, in the best model pair
- 14% → 52%
- GAIA resolve rate from swapping vendor alone, on the same pair
- Task fit
- not forceful wording, drove delegation in a 12-run prompt ablation