This session explores the engineering behind “SCUM: Into the Wild,” a wildlife overhaul that transformed hunting from scripted encounters into a persistent wildlife ecosystem in a live 80-player survival game. The objective was to build a believable world where animals naturally roam, are reliably encountered, and make hunting feel like part of a living ecosystem while operating within the strict performance constraints of dedicated multiplayer servers. The session begins by exploring the limitations of the previous clue-based hunting system and why simply increasing AI counts could never achieve the intended player experience. Instead of designing AI around isolated hunting encounters, the focus shifted to building a living ecosystem. Continuous profiling shaped the architecture throughout development, ensuring every major engineering decision was backed by measured performance rather than assumptions. The presentation examines the shipped architecture, including biome-based population management, persistent wildlife populations, multiple levels of simulation detail, continuous Behaviour Tree execution, movement simulation, perception, networking, AI budgeting, and the profiling and debugging tools used to understand thousands of simultaneously simulated agents. Throughout the session, the architecture is explained through the production challenges that shaped it, highlighting how profiling, experimentation, and engineering trade-offs led to the final shipped solution. Common assumptions about scaling AI systems, including Behaviour Trees and EQS, are examined alongside the real production bottlenecks encountered during development. While “SCUM: Into the Wild” serves as the case study, the engineering principles extend beyond wildlife simulation. The session concludes with practical lessons on simulation relevance, scalable AI architecture, and continuous performance analysis that apply to any game simulating large populations of believable NPCs.