TLDR: Trans-ACT is a new framework that integrates Transactional Analysis (TA) principles, including ego states (Parent, Adult, Child) and life scripts, into LLM-based AI agents. This allows agents to simulate realistic human psychological dynamics and social interactions, as demonstrated in a ‘Stupid game’ simulation, opening avenues for applications in conflict resolution and psychological research.
Artificial intelligence is rapidly advancing, with Large Language Models (LLMs) now capable of understanding and generating human-like text. These LLMs are increasingly being used to create “agents” that can work together in Multi-Agent Systems (MAS) to achieve complex goals. However, traditional AI systems often lack the emotional depth and psychological nuances that characterize real human interactions.
A new research paper introduces Trans-ACT (Transactional Analysis Cognitive Toolkit), a novel approach that aims to bridge this gap by embedding principles from Transactional Analysis (TA) into AI agents. TA, developed by Eric Berne, is a psychological framework that helps understand human behavior and social interaction through concepts like ego states and life scripts. The paper, titled “Games Agents Play: Towards Transactional Analysis in LLM-based Multi-Agent Systems,” was authored by Monika Zamojska and Jarosław A. Chudziak.
Understanding Trans-ACT’s Core
Trans-ACT models an agent’s personality using TA’s three ego states: Parent, Adult, and Child. Each of these states represents a distinct set of feelings and behaviors:
- Parent: Embodies internalized rules, beliefs, and attitudes learned from authority figures.
- Adult: Represents a rational, objective mindset focused on logical information processing.
- Child: Reflects emotional responses and behavioral patterns rooted in early life experiences.
In the Trans-ACT architecture, each ego state is designed as a separate sub-agent. These sub-agents actively search for relevant memories specific to their ego state. For instance, the Parent state retrieves memories of rules, the Adult state accesses factual knowledge, and the Child state recalls emotional responses from childhood. These memories are stored and retrieved using a similarity-based search mechanism, allowing the agents to shape their responses based on past experiences.
After each ego state generates a potential response, a dedicated decision-making agent evaluates these options. This agent considers the relevance of each response, the current social context, and crucially, the agent’s “life script.” A life script, in TA, is an unconscious plan developed in childhood that guides decisions and shapes relationships. By integrating this concept, Trans-ACT ensures that the final chosen response is consistent with the agent’s underlying psychological framework, making interactions more realistic and context-aware.
Simulating Human Psychological Games
To demonstrate Trans-ACT’s effectiveness, the researchers simulated a well-known TA “game” called “Stupid.” In this scenario, one individual (Jordan) pretends to be helpless to gain attention, while another (Alex) assumes a “rescuer” role. The simulation configured Jordan to act helpless and avoid responsibility, believing he needs others to solve his problems. Alex was programmed to fix others’ mistakes to prove his worth.
The agents were given distinct memory structures and life scripts. The simulation used a LangGraph architecture, where each conversational step connects in a graph, allowing for dynamic interactions. The underlying language model for all agent reasoning was GPT-4o, a powerful LLM.
The results showed that agents modeled with TA principles produced nuanced and context-aware responses. Jordan consistently expressed helplessness, while Alex predictably took control, mirroring the “Stupid” game pattern observed in human interactions. The simulation successfully captured variations in emotional tone and demonstrated how memory retrieval shaped real-time decision-making, leading to behaviors consistent with cognitive science principles.
Also Read:
- The Psychological-mechanism Agent: A New Approach to AI Behavior Simulation
- Assessing LLM Therapists: A New Approach to Understanding Client Exploration
Future Directions and Real-World Impact
The integration of Transactional Analysis into Multi-Agent Systems opens up exciting possibilities. Future improvements for Trans-ACT include incorporating reinforcement learning, which would allow agents to learn from experiences and adapt their behaviors over time, potentially even changing their life scripts. Researchers also plan to integrate Trans-ACT with existing cognitive architectures like ACT-R or Soar to enhance reasoning and problem-solving capabilities. Additionally, more advanced TA concepts such as “drivers,” “racket feelings,” and “strokes” could be added to create even more complex and human-like emotional dynamics.
The practical applications of TA-based AI agents are significant. They could be used to develop tools for conflict resolution and mediation, helping to identify and break psychological patterns in disputes. These systems could also advance psychological research by providing models to explore interpersonal dynamics and how early experiences shape behavior. Furthermore, Trans-ACT could enhance scientific research on social systems, simulating how groups with different psychological adaptations respond to social changes or policies, offering valuable insights for social scientists and policymakers. For more details, you can refer to the full research paper here.
