Hillclimbing within this harness

All the typical approaches were still effective
Prompt improvements
Few-shot examples
Making tools better fit-for-use
Breaking the task down
Teaching the agent some expert judgement

43% -> 92%
Accuracy improvement on internal evals