Original title: TRIAGE: Three-level Routing and Intelligent Agent Guidance for Efficient Execution
Ruocan Wei · September 2026
AI agents usually solve problems by thinking through every step from scratch. Even if an agent answers almost the same question twice, it repeats the whole reasoning process. This wastes a lot of computing power and tokens, which are the basic units of text that models process. To solve this, researchers built TRIAGE. It saves past solution paths and turns them into reusable skills so the agent does not repeat work. When a user asks something, TRIAGE checks three tiers. Identical questions get instant answers at zero token cost. Similar questions use an old plan with swapped-in details, also using 0 tokens. Completely new questions run the normal reasoning steps and get saved for next time. In tests, this system cut token use by 62.3% on security queries and 76.3% on general tool tasks. It also improves with use: average token costs fell from 198 to 74.7 within 100 queries.
Imagine you are following a recipe to bake cookies. The first time, you read every step carefully. The next time you make the same cookies, you remember the steps and skip re-reading the recipe. If you want chocolate chip cookies instead of plain, you only adjust the ingredient list, not the whole method. In this paper, TRIAGE works like that memory for AI tasks, storing past steps as templates and using them for similar jobs to save work.
This reduces computational costs for companies and developers using AI agents, making repeated tasks faster and cheaper. It directly cuts down the amount of text processed, which often translates to lower expenses.
The abstract doesn't say how TRIAGE performs on highly dynamic or creative tasks outside the tested domains. It covers experiments on security monitoring and ToolBench, but not universal applicability.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
TRIAGE: Three-level Routing and Intelligent Agent Guidance for Efficient Execution — Ruocan Wei
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.