As AI agents move beyond single-prompt demos into multi-step, autonomous workflows, teams face a core challenge: how do we design, test, and govern agent behavior before deploying it into real systems?

This talk presents simulated worlds as a practical engineering tool for building and evaluating agent systems. Using a controlled, Doom-like simulation environment, we explore how agents can plan actions, coordinate with other agents, operate under explicit policy constraints, and be evaluated over long-running interactions.

Rather than focusing on game mechanics, the simulation is used as a safe, repeatable laboratory for agent engineering. We’ll examine how structured environments help surface failure modes early, make agent behavior observable, and enable meaningful evaluation beyond simple task success.

Topics include:

- Designing agent planning and decision loops in constrained environments
- Enforcing policy and safety boundaries at runtime
- Multi-agent coordination and conflict handling
- Behavioral and lifecycle-based evaluation metrics for agents
- Lessons learned translating simulation insights to real workflows

This session is framework-agnostic and aimed at developers and architects who want reliable, testable, and governable agent systems, not just clever prompts.