TROVE introduces a trace‑grounded route orchestration mechanism that adapts to runtime evidence, addressing the orchestration bottleneck caused by agents committing to execution structures before outcomes are observed. In the offline stage, TROVE distills evaluated workflow‑search traces into atomic skills, composite skills and an outcome‑conditioned transition graph, preserving stable fragments while exposing outcome‑dependent decisions. In the online stage, a planned route is treated as provisional: after committing a top‑level skill, the controller can keep a valid continuation, insert a trace‑supported local response, or replace only the suffix invalidated by evidence. Experiments on code‑generation, question‑answering and math‑reasoning benchmarks with various LLM backbones show that TROVE achieves a stronger quality‑efficiency trade‑off than baselines such as dataset‑level optimization, query‑level architecture selection and graph‑constrained scheduling. Quality gains are largest when outcome changes the appropriate continuation, while early termination yields substantial efficiency gains on near‑saturated tasks. Ablations reveal that composite skills capture most offline benefits, insertion enables local correction, and suffix replacement mainly improves efficiency. These findings establish selective route editing as a general principle for adaptive agent orchestration.
Review