TROVE: Adaptive Agent Skill Orchestration via Trace-Grounded Route Validation and Editing
TROVE is a method for agent skill orchestration that revises only runtime-invalidated parts of a plan. Offline, it distills evaluated workflow-search traces into atomic and composite skills and an outcome-conditioned transition graph. Online, it treats a planned route as provisional, retaining valid continuations, inserting trace-supported local responses, or replacing invalid suffixes. Evaluation across code-generation, question-answering, and math reasoning benchmarks with different LLM backbones shows a stronger quality-efficiency trade-off than existing baselines.
TROVE (Trace-grounded Route Orchestration via Validation and Editing) addresses the orchestration bottleneck in agent systems where pre-execution commitment to a plan can lead to stale steps or broad replanning when intermediate evidence invalidates the continuation. The method revises only what runtime evidence invalidates. Offline, it distills evaluated workflow-search traces into atomic and composite skills and an outcome-conditioned transition graph, preserving stable fragments while exposing outcome-dependent decisions. Online, it treats a planned route as provisional: after committing one top-level skill, the controller retains a valid continuation, inserts a trace-supported local response, or replaces only the invalid suffix. Evaluation across code-generation, question-answering, and math reasoning benchmarks with different LLM backbones shows that TROVE delivers a stronger quality-efficiency trade-off than existing baselines of dataset-level optimization, query-level architecture search, and related methods.
TROVE introduces trace-grounded route validation and editing, using an outcome-conditioned transition graph distilled from workflow-search traces. This allows local, evidence-driven revision of agent execution plans rather than full replanning, improving quality-efficiency trade-offs across diverse benchmarks and LLM backbones.
The approach targets a key limitation in agent orchestration: the cost and error compounding from rigid pre-execution planning. By enabling adaptive, local revisions, TROVE could reduce compute waste and improve reliability in agentic workflows, making it relevant for enterprises deploying multi-step AI agents.
TROVE offers a potential efficiency gain for agent-based products by reducing unnecessary replanning and preserving valid execution progress. This could lower inference costs and improve task success rates, benefiting companies building autonomous agents for coding, QA, and reasoning applications.
Observable next signals include adoption of TROVE-like methods in agent frameworks, further benchmarks on real-world agent tasks, and comparisons with emerging adaptive planning techniques. Potential extensions may address multi-agent coordination and integration with tool-use systems.