Building a Doom-Like World to Explore Agentic Systems - Alexander Chernov - NDC Toronto 2026

Value of the Information & Strength of the Argument

The talk provides valuable insights into using game engines as testbeds for agentic systems, offering a concrete architecture with practical implementation details. The argumentation is strong, supported by a live demo and clear explanations of design decisions. The speaker effectively argues that game engines offer unique advantages for testing agent behavior, such as deterministic loops and shared state. However, the talk is more of an expert opinion and demonstration than a rigorous empirical study, lacking quantitative evaluation of the system’s performance or comparison with alternative approaches.

Scientific Rigor, Source Quality, Title Accuracy

The talk is based on the speaker’s personal experience and does not cite external sources. The title accurately reflects the content. The speaker mentions using GZDoom and Project Brutality but clarifies he is not affiliated. The talk is a conference presentation, so it is not peer-reviewed, but it provides a detailed technical account. The description includes links to NDC conferences, which are not directly related to the content.

The title accurately reflects the content, which focuses on building a Doom-like environment to explore agentic systems.

7/10

The talk presents a detailed architecture for using a game engine as a testbed for agentic systems, with concrete implementation details and a live demo. However, it is based on personal experience and lacks formal evaluation or peer review.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction to the talk and the concept of agents as semantic mirror of the game world.
• 05:00Explanation of why game engines are suitable for agent work: control loops, states, real-time feedback.
• 10:00Overview of the three-tier architecture: structural, semantic, and behavioral mirrors.
• 15:00Detailed discussion of world state modeling, including spatial, zone, agent, policy, and scenario states.
• 20:00Explanation of agent perception with two-tier vision: cheap observer swarm and LM broker swarm.
• 25:00Live demo showing agents identifying objects and reacting to gameplay.
• 30:00Discussion of the control plane, including 1000 agents, brokers, and semantic cache.
• 35:00Explanation of decision loops, FSM states, roles, and actions.
• 40:00Details on determinism, autonomy, and observability of agent actions.
• 45:00System overview: three-layer architecture with game runtime, control plane, and MCP gateway.

• NDC Conferences — Conference organizer, mentioned in description.
• NDC Toronto — Conference website, mentioned in description.

• Model Context Protocol — MCP is used in the talk for external control and integration.