Halo AI Pentest Tool to Split Single Agent Into Five Specialized Roles

Halo, an AI-powered penetration testing system, is being redesigned to replace its single-agent architecture with a five-agent orchestration model. The current setup uses one loop and one shared context window to handle reconnaissance, exploitation, and self-correction, which has led to increasingly muddled reasoning over long engagements. The planned architecture introduces five distinct agents — Planner, Orchestrator, Vulnerability Discovery, Attacker, and Debugger — each handling a focused subset of the overall pentest workflow. Separating these responsibilities is intended to reduce context pollution, make failures easier to isolate, and prevent errors in one phase from cascading into others. The multi-agent design is still in early sketching stages, with details on inter-agent communication and shared state yet to be finalized.