Skip to content

AI Players

AI Players are autonomous LLM-powered agents that play MAID as virtual players. They connect through virtual sessions, perceive the world through game output, and act by issuing commands — exactly like human players, but driven by a cognitive architecture.

Unlike NPC dialogue (which gives NPCs AI-powered conversation), AI Players are fully autonomous agents with their own goals, memories, and decision-making. They explore, fight, trade, form groups, and even file bug reports when they encounter game issues.

Key Concepts

  • Virtual Sessions — AI Players plug into the existing Session protocol. The game engine can't tell them apart from human players.
  • Three-Layer Architecture — A hybrid control system with fast reactive reflexes (zero LLM), a mid-speed executive loop (cheap LLM), and slow deliberative planning (expensive LLM). The fast loop never waits for the slow loop.
  • Memory-Driven Learning — Five memory layers (working, episodic, semantic, procedural, reflective) let AI Players learn from experience, remember locations, and build reusable command sequences.
  • Personality System — Big Five personality dimensions (openness, conscientiousness, extraversion, agreeableness, neuroticism) shape behavior: how aggressively they fight, whether they greet strangers, how quickly they flee.
  • Cost-Managed LLM Usage — A tiered model strategy keeps costs around $0.10/agent/hour. Budget enforcement degrades gracefully — reflexes keep working even when the LLM budget runs out.
  • Bug Filing — AI Players detect game anomalies (broken exits, command errors, state inconsistencies) and file structured, deduplicated bug reports.

How They Work

Game World ──→ Perception ──→ Memory ──→ Planning ──→ Action ──→ Game World
                                ↑                        │
                          Reflection ◄───────────────────┘
                        (Layer 3, async)

Each cognitive tick, the AI Player reads game output, updates its world model, checks its plan, and issues a command. Periodically, a background task reflects on accumulated experience and revises long-term goals.

See Also