Most agent orchestration methods fix a workflow before execution, so when intermediate results make the next step unnecessary or wrong, the agent either follows a stale path or replans from scratch. TROVE instead treats a planned route as provisional. It commits to one top-level skill, observes the outcome, and then keeps the valid continuation, inserts a local corrective step, or replaces only the invalid suffix.
Offline, TROVE distills evaluated workflow-search traces into atomic and composite skills plus an outcome-conditioned transition graph, preserving stable fragments while exposing outcome-dependent choices. Online, this lets the controller edit the route at the smallest scope that the new evidence requires.
Across six code, question-answering, and mathematics benchmarks with three LLMs, TROVE achieved the best or tied-best task score in 15 of 18 settings and the shortest online time in 16. Relative to AFlow, it improved task score in 16 settings by up to 31.37% and reduced time by up to 86.7%. The gains are largest when outcomes change the appropriate continuation, and early termination gives large efficiency gains on near-saturated tasks.