LLM Agent Routing
LLM: Large Language Model
Overview
LLM agent routing studies how to match tasks and changing execution states with suitable models, tools, skills, or groups of agents. Common approaches include retrieval and reranking over capability descriptions, learned classifiers and contextual bandits, reinforcement-learning orchestrators, and confidence- or progress-based escalation; some systems also impose workflow, budget, or safety constraints. Evaluation spans tool use, coding, question answering, and long-horizon interactive tasks, with growing attention to end-to-end success, latency, token cost, and recovery from failures. The central challenge is showing that routing reliably beats a strong fixed choice: gains can be sensitive to noisy feedback, changing task distributions, large or deceptive capability catalogs, and the difficulty of evaluating stateful handoffs.
Latest papers
99 papers16 papers in the last four weeks, up 14% on the four weeks before. 0.1% of all new papers.