How Agents Work With Parallel Agents

Rémy Duthu discusses his evolving approach to using multiple parallel AI coding agents, highlighting how his initial limitation of two agents due to cognitive load has shifted to managing a dozen with improved efficiency. Originally, the challenge was the intense monitoring required for each agent, which multiplied with the addition of more agents. However, advancements in agent autonomy reduced the need for constant supervision, enabling him to run more agents simultaneously. The real shift came from recognizing and addressing the cognitive load associated with tracking and retrieving specific agent sessions, which was more burdensome than the mere number of agents. By implementing a system of session renaming and pinning, Duthu streamlined monitoring and retrieval processes, effectively managing up to ten sessions concurrently. Despite these improvements, the increased volume of tasks completed by agents has shifted the bottleneck to the review process, necessitating new strategies for managing code reviews, such as breaking down changes into smaller, more manageable pieces and relying on a layer of agentic reviewers before human intervention. Duthu acknowledges the shift from writing code to orchestrating agents and is actively exploring the impact on review workflows to ensure quality and efficiency.