TRIAGE addresses the inefficiency of ReAct-based LLM agents that repeat full reasoning loops for similar queries. It introduces a three-level routing system: identical queries are reused directly, similar queries are handled by substituting parameters in a cached trajectory, and novel queries are executed with full ReAct and stored for future use. A semantic encoder retrieves matching trajectories, and an automatic skill extraction mechanism distills frequent patterns into deterministic skills.
Experiments on 1,007 security monitoring queries achieved 62.3% token savings, with over 60% of queries executed at zero token cost. Cross-domain validation on 345 ToolBench queries showed 76.3% token reduction, demonstrating generalizability. An online learning experiment revealed a cold start where skill hits rose from 0% to 57% within the first 100 queries, and average token cost dropped from 198 to 74.7. The framework creates a positive feedback loop where usage improves efficiency.