How AI Agents Repeat Tasks: Prompts, Skills, Graphs, or Code

This Habr article explores the challenge of task reproducibility in AI agent systems. The core issue is that agents often re-plan steps upon re-execution, consuming extra tokens and potentially producing inconsistent output formats. The author examines four approaches to solve this: saved prompts, agent skills, execution graphs (workflows), and code-based action caching. The article details the pros and cons of each method, covering topics such as model non-determinism at zero temperature and error handling during mid-process failures. It also addresses strategies for preventing unintended side effects during repeated runs. This analysis is highly relevant for developers building agentic systems who need to improve the stability and predictability of their workflows.