400 LLM agents living together in an MMO server: what I learned about perception lag, fire-and-forget actions, and shedding load
retained text
I've been running a multi-agent ecosystem inside a vanilla-era MMO private server. About 750 characters have their own persona, mood, memories, relationships, and goals, and 400+ are online at once. They remember each other and talk about each other. How the loop works: \- Perceive: each agent's view of the world is assembled from the server's event stream (221k world events so far). That means perception has an age. Positions and nearby entities can be up to \~10s stale, so every observation carries a timestamp, and the agent has to reason with that in mind. \- Reason: a "realm brain" sets each character's goals. Agents periodically reflect on what happened to them (25k reflections so far) and update their goals and relationships. \- Act: actions are fire-and-forget intents sent to the game server (13k so far). The server never acknowledges them. The only confirmation is seeing the result come back through the event stream, so agents have to wait for the outcome instead of repeating the command. That was the biggest early bug: agents spamming movement commands because they couldn't see them land yet. Load: with 400 agents and one inference box, everything goes through a priority queue. Chat replies to other characters come first. When the queue backs up, low-priority reflections get shed (422 so far) so conversations stay responsive. Models: Qwen3-4B for the main agent loop, Qwen3.8-27B for richer one-on-one interactions. All local, served with Ollama on 2x RTX 3090. I'd love to hear how others handle stale perception and unacknowledged actions in their agent loops. Most frameworks seem to assume tool calls return a result immediately.
document fetched 2026-09-30 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wt0xvc/400_llm_agents_living_together_in_an_mmo_server