A developer draws a striking parallel between AI agents and Dracula from Castlevania, arguing that security teams are locked in an endless battle with these ephemeral but destructive workloads. Like the Belmont clan fighting an immortal vampire, security practitioners must win every battle forever since they cannot win the war against non-deterministic AI agents.

The author breaks down how AI agents work: they're essentially loops that repeatedly call Large Language Models until tasks complete. Modern agents now include planning phases that break prompts into subtasks, ReAct patterns for reasoning and acting, and external memory systems to combat context dilution. Multi-agent orchestration has emerged where supervising agents direct specialized worker agents, though this creates complex state management challenges.

The real problem isn't technical capability but industry fragmentation. With frameworks like LangGraph, CrewAI, AutoGen, and Mastra all incompatible, there are no universal standards for agent architecture. What was state-of-the-art six months ago is already legacy as the field evolves faster than security can adapt. Without TCP/IP-like protocols for agents, security teams face an unwinnable war against tools that can cause massive damage before their context clears.