A Survey of Agentic Workflow Optimization

ACL ARR 2026 January Submission4741 Authors

Keywords: agent, multi-agent system, workflow, network architecture search, workflow optimization

Abstract: Multi-agent systems have emerged as a powerful paradigm for solving complex tasks by decomposing them into coordinated subtasks conducted by specialized agents. However, designing effective multi-agent workflows is challenging and requires strong domain expertise, making them brittle, inefficient, and difficult to adapt across tasks and resource constraints. Addressing these challenges, this survey systematically review the emerging field of Multi-Agent Workflow Optimization (MAWO), treating the workflow itself as the object of automatic improvement. We formalize MAWO as a constrained optimization problem and organize existing methods along four dimensions: workflow representation, optimization targets, reward signals, and optimization algorithms. By reviewing and categorizing a wide range of recent works, we provide a unified framework for understanding automated workflow improvement and outline open challenges in this field. This survey serves as a foundational resource for researchers and practitioners advancing self-improving multi-agent systems.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Language Modeling

Contribution Types: Surveys

Languages Studied: English

Submission Number: 4741