400 LLM Agents Living in an MMO Server: Lessons on Stale Perception, Fire-and-Forget Actions, and Load Shedding
kristiantalley679 · reddit · 2026-09-29
A developer ran a multi-agent ecosystem inside a vanilla-era MMO private server: 750 characters each with persona, mood, memories, relationships, and goals, 400+ online simultaneously, logging 221k world events, 25k reflections, and 13k actions.
Key architecture insights:
- Perceive: each agent's worldview is assembled from the server event stream; positions can be 10s stale, so every observation carries a timestamp and agents must reason about data age.
- Reason: a "realm brain" sets goals; agents periodically reflect and update goals/relationships.
- Act: actions are fire-and-forget intents with no server acknowledgment—confirmation only comes from seeing results in the event stream. The biggest early bug was agents spamming movement commands they couldn't see land.
Load: with 400 agents on one inference box, a priority queue prioritizes chat replies and sheds low-priority reflections (422 shed) under backlog.
Models: Qwen3-4B for the main loop, a 27B model for rich one-on-one interactions, all local via Ollama on 2x RTX 3090. The author notes most agent frameworks assume tool calls return immediately, unlike game servers.
Related event: 400+ LLM Agents Live Together in a Classic MMO Server(2 posts)→
More from AGI Musings
- Pedro Domingos: neolabs could be key to AGI — if they don't sell out to big companies — pmddomingos · 2026-09-29
- Reply to Bengio: the real explosion is throughput, not intelligence, widening the audit gap — AryHHAry · 2026-09-29
- Two recursive mechanisms for an intelligence explosion: automated AI R&D expands the effective workforce — jeffclune · 2026-09-29
- Heavy AI remixing erodes attribution: misattribution feels worse than copying, researcher warns — gandamu_ml · 2026-09-29
- Bosses with 'AI brain': 100 marketers say ideas now wait on Claude's verdict, and a product gap emerges — cen6wkf · 2026-09-29
- Dev argues AI harms trace to human intent or incompetence — and AI itself is the fix — davidpattersonx · 2026-09-29