Topic

From Complexity to Breakthrough: Practical Implementation of Coding Agents and Multi-Agent Principles

In algorithm research, repetitive tasks such as script writing, unit test generation, and code exploration have long consumed researchers’ time and energy. Coding Agent, developed by Baidu’s Comate team, significantly improves efficiency through natural language–driven workflow automation, automated testing, and codebase Q\&A. More importantly, it introduces multi-agent collaboration at a fundamental level: an Architect Agent decomposes complex tasks into subtasks, delegates them to specialized agents for parallel execution, and integrates their outputs into a complete solution. This talk will showcase real-world use cases demonstrating how Coding Agent enables rapid paper reproduction, cross-repository comprehension, and quality assurance. It will also provide a deep dive into the design principles and practical experience behind its multi-agent coordination mechanisms, offering attendees actionable insights and a clear path to applying these ideas in their own work.