TRIAGE: Three-level Routing and Intelligent Agent Guidance for Efficient Execution
TRIAGE claims zero-token execution for repeated or similar agent queries by reusing prior trajectories as skills.
The paper proposes a three-level router for ReAct-style LLM agents: direct reuse, deterministic skill substitution, or a full new ReAct run. In experiments on 1,007 security monitoring queries, it reports 62.3% token savings, with 61.5% of queries handled at zero cost across the first two levels. A ToolBench validation across 15 domains reports a 76.3% token reduction. The system also stores novel executions and extracts high-frequency patterns into reusable skills over time. ArXiv · AI/CL/LG's note
The paper proposes a three-level router for ReAct-style LLM agents: direct reuse, deterministic skill substitution, or a full new ReAct run. In experiments on 1,007 security monitoring queries, it reports 62.3% token savings, with 61.5% of queries handled at zero cost across the first two levels. A ToolBench validation across 15 domains reports a 76.3% token reduction. The system also stores novel executions and extracts high-frequency patterns into reusable skills over time. ArXiv · AI/CL/LG's note
score 5