Why AI Agents Drift Mid-Task and How a Multi-Agent System Fixes It

This post introduces a multi-agent system architecture for AI agents to prevent task drift. It details the problem of context window limitations and proposes a solution where complex tasks are broken down into specialized agents (Code Cartographer, Engineering Alchemist, Visual Stylist, Ticket Auditor, Review Agents). The core innovation is anchoring these agents to Jira, using issue links as a dependency DAG and work items as prompts. A deterministic daemon orchestrates agent execution based on Jira ticket status, ensuring reliable scheduling and state management, which is contrasted with less stable local markdown file approaches. The post outlines a four-step process for building such a system: anchoring to Jira, running a watching daemon, giving agents personas, and reporting back to advance the loop.