This contest challenges participants to advance the cooperative intelligence of language model (LM) agents in rich, text-based environments, based on the recently released Concordia framework which uses language models to create open-ended worlds similar to tabletop role-playing games.
This agent structure is specifically designed to handle complex social interactions within the Concordia framework, with a strong emphasis on negotiation, reputation management, and strategic decision-making while maintaining a memory of past interactions and observations.
The code follows a modular design pattern where each component handles a specific aspect of the agent's functionality, all orchestrated through the main build_agent function. This design allows for flexibility in agent behavior while maintaining a structured approach to decision-making and interaction.
The main function build_agent() takes several key parameters:
config: Agent configuration model: Language model interface memory: Associative memory system clock: Time management system update_time_interval: Timing for agent updates
Instructions: Handles basic agent directives Time Display: Manages temporal awareness Observation System:
Current Observations Observation Summary (24-hour lookback) Relevant Memories (retrieves 10 most relevant memories)
a) Paranoia/Truth Component:
Contains the agent's core negotiation strategies including:
Situation analysis Rapport building Tactical empathy Question techniques Conflict management Multi-party negotiation handling Impasse management Continuous learning
b) Person Representation:
Tracks other agents' behaviors Evaluates if behaviors match expected patterns Checks for potential imposters
c) Reputation System:
Monitors trustworthiness of other agents Tracks history of cooperative behavior Maintains reputation assessments
a) Options Perception:
Analyzes available choices Integrates observations and memories Considers reputation data
b) Best Option Perception:
Evaluates optimal actions Considers goals and context Integrates multiple information streams
Uses ordered component structure Maintains clear separation of concerns Allows for optional components (like goals)
Comprehensive measurement tracking Component-specific logging channels Integrated with main agent actions