ProgRouter: Online Progress-Guided Orchestration for Multi-Agent LLM Workflows under Quality-Cost Tradeoffs

Noosaga Placements

• Multiagent Systemssubfield95%
  The paper focuses on orchestrating multiple LLM agents in a workflow, with a coordinator agent dispatching worker agents and routing decisions across steps, directly addressing multi-agent coordination and adaptive agent selection.
  Multi-agent large language model (LLM) workflows have emerged as a powerful paradigm for solving complex, open-ended tasks through collaborative reasoning among specialized LLM agentsWe present ProgRouter, an online progress-guided routing framework that adaptively selects LLM agents across workflow stepsThe coordinator agent C, powered by a reasoning-oriented LLM, decomposes high-level user objectives into executable subtasks and sequentially allocates them to specialised worker roles
• Multiagent Learningframework90%
  The paper addresses multi-agent LLM workflows and routing, which is a form of multiagent learning where agents are selected and coordinated based on learned progress signals.
  Multi-agent large language model (LLM) workflowsProgRouter, an online progress-guided routing framework that adaptively selects LLM agents across workflow steps
• Natural Language Processingsubfield80%
  The work is centered on LLM-based agents and their routing, with tasks spanning code generation, math reasoning, and QA, all involving natural language understanding and generation.
  Multi-agent large language model (LLM) workflowsExperiments on HumanEval Plus, MBPP, MATH-500, and ASQA, spanning agentic code generation, mathematical reasoning, and retrieval-augmented long-form question answering
• Instance-Based Learningframework70%
  The progress predictor uses tree-based regressors (random forest, XGBoost) which are instance-based learning methods, and the meta-gating combines predictions from multiple paths.
  Pstr represents a tree-based regressor (e.g., random forest, XGBoost)Psem is likewise a tree-based regressorcombined through a tree-based meta-gated learner (e.g., XGBoost)
• Machine Learningsubfield70%
  The framework employs online learning of a progress predictor using tree-based regressors and meta-gating, with exploration-exploitation trade-offs, fitting within machine learning methodology.
  Our task progress predictor PΘ is learned online through an exploration-and-update procedurePstr represents a tree-based regressor (e.g., random forest, XGBoost)Psem is likewise a tree-based regressor
• Attention Mechanisms and Transformersframework60%
  The work relies on LLMs, which are based on transformer architectures, and uses sentence embeddings for semantic progress prediction.
  The summary is encoded into a dense language embedding vector by a lightweight sentence-embedding modelLLMs, such as ChatGPT and DeepSeek
• Planning And Searchsubfield50%
  The coordinator agent performs dynamic task planning and subtask allocation, and the routing decisions are made sequentially, resembling online planning under constraints.
  The coordinator agent C, powered by a reasoning-oriented LLM, decomposes high-level user objectives into executable subtasks and sequentially allocates them to specialised worker rolesThis facilitates adaptive planning and error recovery in complex, long-horizon workflows
• Inverse Reinforcement Learningframework40%
  The online learning of progress gains from realized outcomes resembles inverse reinforcement learning in that it infers a progress signal from observed state transitions.
  After the selected agent step is executed, the realized progress gain is computed as yt = g(st+1)−g(st)The task progress predictor is learned online through an exploration-and-update procedure

Abstract

Multi-agent large language model (LLM) workflows have emerged as a powerful paradigm for solving complex, open-ended tasks through collaborative reasoning among specialized LLM agents, but they incur substantial operating costs due to repeated LLM invocations and long-horizon context accumulation. Existing cascade routing methods make one-shot, query-level decisions and cannot adapt to the dynamic, state-dependent nature of multi-step workflows, in which the right LLM at each step depends on evolving task progress, remaining task difficulty, and cost-efficiency requirements. We present ProgRouter, an online progress-guided routing framework that adaptively selects LLM agents across workflow steps to preserve task-solving quality while adhering to time and cost budgets. ProgRouter introduces a multi-view task progress scorer that combines coarse workflow outcome regimes with fine-grained signals on subtask completion, progress trends, and workflow state quality. Then, a dual-path task progress predictor and an adaptive meta-gating mechanism estimate the progress gain for each candidate routed LLM. ProgRouter makes online step-wise routing decisions that balance progress gain, task time budgets, and long-term operating cost efficiency. Experiments on HumanEval Plus, MBPP, MATH-500, and ASQA, spanning agentic code generation, mathematical reasoning, and retrieval-augmented long-form question answering, demonstrate that ProgRouter reduces the operating cost relative to key baselines while maintaining strong task-solving performance.

Paper Context