Behavior is all you need

Aug 2025

Authors:

Vishnu Hari
Connor Brennan

Abstract

Current game bots lack the emotional depth, spontaneity, and intent that make human players compelling, resulting in predictable interactions and broken immersion. We propose a framework for building believable AI agents that behave convincingly enough to evoke emotional engagement without requiring true sentience. Drawing from narrative theory and human psychology, our approach focuses on four key components: personality, needs, emotion, and memory. Together, these modules drive behavior that appears motivated, context-sensitive, and consistent over time. Much like how audiences connect with fictional characters, players can suspend disbelief when bots behave in ways that suggest underlying intention. This play-acting approach offers a scalable path to lifelike game agents - enabling emergent gameplay, dynamic world population, automated testing, and emotionally resonant interactions - not by replicating consciousness, but by engineering behavior that is rich enough, reactive enough, and real enough to be believed.

Introduction

Humans make multiplayer games compelling because of unpredictability and emotionally meaningful choices. By contrast, most bots are deterministic, stateless, and emotionally flat. The whitepaper observes that players don't need literal minds from NPCs—they need behavior that suggests intention. Much like audiences invest in fictional characters, players will suspend disbelief when agents act with coherent motivations and emotional modulation.

Framework in Brief

The paper proposes a practical behavioral stack that yields believable agents without requiring full cognitive realism.

Personality

Long-term tendencies and style. Personality shapes how an agent interprets events and chooses among plausible actions so behavior feels consistent over time.

Needs

Explicit drives (e.g., belonging, curiosity, success) create immediate motivation and make behavior legible to players.

Emotion

Transient affect that modulates decisions and responses to praise, conflict, risk, and reward— crucial for making agents feel alive.

Memory

Short- and long-term memory preserve context and growth, preventing stateless resets that break believability.

System Architecture

A lightweight, interpretable controller surrounds a language model core. Modules inject state (personality, needs, emotion, memory) and the agent acts via a tool/action interface bound to world state.

Whitepaper figure: high-level architecture — Figure 1. High-level agent architecture.

Control loop

Sense world + retrieve relevant memory
Condition policy on personality, needs, and emotion
Select action via tools or language; update internal state

Tool/action interface

Grounded actions and tool calls enable coherent non-language behavior alongside dialogue.

Data collection and model path

With sufficient high-quality gameplay data, the controller can be consolidated into an end-to-end foundation model that internalizes these dynamics. Training blends supervised behavioral cloning, RLHF, self-play, and simulated data bootstrapping.

Evaluation: Embodied Turing Test

The paper proposes a multiplayer deception task where humans attempt to identify AI-controlled players using linguistic, strategic, and movement cues. Success is defined by near-chance human detection, alongside secondary metrics like role alignment, strategic diversity, and adaptive style shifts.

Applications and benefits

More authentic gameplay and emergent social dynamics
Automatic population of worlds with believable agents
Automated play/balance testing and QA at scale

Limitations and future work

Latency and cost constraints for real-time control
Alignment of personality/needs with game design intent
Memory scaling and privacy considerations

Conclusion

The core insight is pragmatic: agents do not need true minds to feel human—they need consistently motivated, emotionally expressive behavior. By centering design on the appearance of intention and continuity, we unlock richer gameplay, emergent social dynamics, and scalable world population.

For the full details, see the original whitepaper: Behavior is all you need