Large language model (LLM) agents are increasingly applied to software engineering tasks, yet their development and maintenance remain challenging, particularly due to di!culties in understanding agent behavior. Developers must reason about trajectories of LLM queries, tool calls, and code modifications, but current tools reveal little of this intermediate process in a comprehensible format. Consequently, diagnosing failures and understanding agent behavior remain major obstacles. Drawing on parallels with conventional software development, we introduce AgentStepper, an interactive debugging approach for LLM agents that adapts established debugging practices to agent trajectories. AgentStepper records trajectories as structured conversations among the LLM, the agent program, and tools. It supports breakpoints, stepwise execution, and live editing of parameters, while automatically capturing and linking repository-level code changes. The interface follows familiar paradigms from chat applications, debuggers, and IDEs. Our evaluation shows that AgentStepper integrates with representative agents with modest e”ort and improves developers’ ability to interpret trajectories and identify bugs, thereby reducing workload and frustration compared to conventional tools.