决策时刻指导：保持 Replit 智能代理的可靠性

This article discusses the challenges and solutions in developing Replit's intelligent coding agent system. As the agent tackles more complex tasks, session durations have increased and trajectories have grown longer, leading to risks of compounding model failures and unexpected behaviors. Traditional static prompt rules often fail to generalize or may contaminate the context when scaled. The paper proposes a novel approach: utilizing the execution environment itself as a guidance mechanism. By enabling the environment to provide intelligent feedback, it helps the agent course-correct while keeping humans in the loop. These techniques have proven effective on long trajectories, significantly improving Replit Agent's performance in building, planning, deployment, and overall code quality, while maintaining reasonable costs and context management.