Halo, a system designed for pentesting, is evolving from a single-agent model to a multi-agent orchestration to improve its functionality and address limitations. Previously, one agent managed all aspects of an engagement, leading to cluttered context, muddled reasoning, and compounded errors. A pentest involves distinct phases like planning, execution, and self-correction, which are currently consolidated in Halo's single-loop system. Separating these tasks into specialized agents will streamline processes, allowing each agent a focused job and an uncluttered context window. This also aids in isolating and addressing failures more effectively.The new architecture will feature five distinct agents. The Planner will break down high-level engagement goals into objectives, focusing on strategy. The Orchestrator will manage agent interactions, routing information and deciding action sequences. The Vulnerability Discovery agent will identify and prioritize weaknesses based on scan data. The Attacker will execute specific tasks against defined targets. Finally, the Debugger will monitor for failures, diagnose issues, and feed insights back into the system to prevent repetition. This design prioritizes defining roles and boundaries before coding, preventing costly reworks. The multi-agent approach is crucial for Halo's scalability, enabling it to run longer, recover gracefully from errors, and improve autonomously over time. This architectural shift marks a significant milestone for Halo's development.