Researchers introduce ActiveSaddler, a framework that treats the training curriculum for LLM agent harness optimization as a non-stationary bandit problem rather than a fixed scenario order. The system abstracts recurring agent failures into reusable failure-pattern arms and adaptively balances revisiting known weaknesses against exploring new ones as the harness evolves. On the GAIA2 and Terminal-Bench 2.0 benchmarks, ActiveSaddler improved test Pass@1 by 4.4 and 7.5 percentage points respectively over the same harness optimizer using a scenario order fixed before optimization.
