Benevolent AI Agent Interactions

How Our Project Will Contribute To The Growth Of The Decentralized AI Platform

As agents and agentic systems are the latest industry-wide AI paradigms, with their design patterns incorporated into mass-consumer applications (i.e., chatbots), having a framework that enhances healthy human-agent interactions reduces toxicity and greatly improves the general online well-being and hygiene. This is greatly in line with the BGI mission of developing beneficial AI, especially around instilling human qualities of fundamental/critical technologies like AI agents.

Our Team

Our team is composed of the following members:

• Kevin R.C.: Project owner, lead data scientist, ML research engineer
• Daniel S.: Principal research scientist and NLP expert NLP
• Chibudem B..: Software engineer and UX research
• CJ L.: AI engineer and library builder

Our team is highly experienced in AI research and service implementation, with a lengthy track history including being awarded for the following services:

• SYBIL (DF-R2)
• Onboard NeuralProphet (DF-R2)
• MAUQ (DF-R3)
• Deep SYBIL (DF-R4)

Company Name (if applicable)

The core problem we are aiming to solve

The internet and social media can be a toxic place, with harassments, brigadings, and scams abound. As AI agents become more prevalent and human-agent interactions become widely normalized, they can inherit and even multiply this toxicity without proper human-centric guardrails or processes being put in place. Recent GenAI and LLM innovations are focused on improved complex reasoning and logical problem-solving capabilities while reducing computing costs to make them widely accessible (i.e. DeepSeek). However, more priority can be placed on human-centric interactions with agents, making them more wholesome and engaging and less sterile and even harmful.

Our specific solution to this problem

We will develop an open-sourced Python library or SDK that provides a framework for agents to interact more benevolently with humans or even other agents. Any developer can use this library to incorporate their own AI agents.

Our library may contain the following features (and GenAI tools):

1. Sentiment analysis (NLTK, Spacy, Langchain)

2. Agent memory/state bank parameters that tune the emotions

• Personality and emotion parameters (Langchain, Langgraph)

3. Chain-of-Emotion (CoE) prompts

• Variant of Chain-of-Thought prompts (Langchain, Langgraph)

4. Other cognitive or logic-based prompting or retrieval methods outside of LLMs

• Neural-Symbolic (PyNeuralLogic)

• Category Theory (Pseudocode written in the prompts)

• OpenCog Hyperon

• Other SNET services (MeTTa KG or MAGUS)

After building this library, we will experiment with the framework with our own McDolan agent or other character agents to see which features, or a combination of features, will work best at which interaction points. We will detail the framework architecture and the experiment results in a final whitepaper, which we plan to present at an upcoming AGI or BGI conference or other related events.

Project details

Our framework will leverage a combination of advanced AI techniques, sentiment analysis, and personality modeling to create a dynamic and adaptive conversational system. Although not set in stone, our framework will likely focus on the following key components:

1) Sentiment Analysis and Personality Factors

• Sentiment Analysis: We will incorporate sentiment analysis to detect the emotional state of the user (e.g., positive, negative, or neutral). This will allow the system to tailor its responses dynamically based on the user’s current mood.

• Personality Modeling: We will integrate established psychological frameworks such as the Big Five Personality Traits (openness, conscientiousness, extraversion, agreeableness, neuroticism) and MBTI (Myers-Briggs Type Indicator) to simulate diverse user personas. For example, we can manually configure a user profile with high neuroticism or low agreeableness to test how the system adapts to challenging interactions.

2) Agent Persistent Memory State that Stores Emotion Parameters

The system is designed to dynamically generate and augment prompts based on the agent's emotional state, leveraging a memory mechanism to provide additional context. Emotional states (e.g., patience, anger, joyfulness) are used as keys to retrieve relevant prompt keywords and/or context vectors from external persistent memory storage. These keywords and/or context vectors are then used to sculpt and enrich the prompts, ensuring that the responses are personalized, context-aware, and emotionally appropriate.  They will then be fed into a chain-of-emotion (CoE) prompt template, which we discuss in the next section.

3) Chain-of-Emotion (CoE) Prompting

Chain-of-Emotion (CoE) is a variation of Chain-of-Thought (CoT) prompting, but instead of prompting the LLM's thought process, you prompt in what is the emotional state.

For example, one of the popular CoT prompt templates is ReAct, which prompts LLM to spell out what is its thought, action, observation.  Here is an example ReAct template from LangSmith.

In the CoE variation of the ReAct template, we can replace the thought component of the prompt with the emotion, specifying its emotional states. This is where we will also plug in the emotional keywords or context vectors as set by the emotion parameters in the agent persistent memory space as mentioned in 2). We will further refine and develop multiple variations of the CoE template, as well as how to best plug-in the emotion parameter values into it.

4) Other Cognitive or Logic-Based Methods

The sentiment analysis, personality modeling, and Chain-of-Emotion (CoE) prompting are more from the NLP+LLM branch of AI. In addition to those, we explore other cognitive and logic-based methods that can enhance the effectiveness of bot-to-bot communication, particularly in identifying, escalating, or resolving conflicts. Two such methods include PyNeuralLogic and Category Theory, which can be integrated into the prompt engineering process to improve reasoning, adaptability, and conflict mitigation strategies.

PyNeuralLogic for Logical Inference in Conflict Detection and Resolution

PyNeuralLogic is a framework that integrates symbolic reasoning with neural networks, allowing for robust logical inference within conversational AI. By embedding logical rules into the prompt structure, we can ensure consistency and coherence in bot interactions, particularly in detecting emotional escalation and suggesting de-escalation strategies.

Category Theory for Structural Understanding in Multi-Agent Emotional Dynamics

Category Theory provides a mathematical foundation for understanding relationships between abstract concepts. By applying category-theoretic structures to prompts, we can design a more structured approach to bot communication, ensuring logical flow, emotional consistency, and adaptability in dynamic emotional exchanges.

Incorporation of OpenCog Hyperon-Based Projects

We also plan to incorporate projects from the OpenCog Hyperon ecosystem, such as Scalable MeTTa Knowledge Graphs and Modular Adaptive Goal and Utility System (MAGUS). We can potentially incorporate them into our framework once their library or AI service becomes available.

Integration with Existing Methods

By incorporating PyNeuralLogic and Category Theory into our existing framework, we create a more robust, logically consistent conversational system that can:

• Detect potential conflicts early using logical rules and neural-symbolic reasoning.
• Ensure structured, mathematically rigorous emotional transitions in conversations.
• Adaptively mediate and resolve conflicts in multi-agent interactions.

Through these cognitive and logic-based methods, our project aims to refine bot-to-bot communication, making it more context-aware, self-correcting, and effective in managing emotionally charged interactions.

-----------------------------------------------------------------------------------------------------

While developing and demoing this framework, we will use McDolan and/or our other AI agents to test this framework. We will be like the alpha users of this framework, where we import the library within our internal McDolan code base to test how the library's functionalities will affect McDolan's interactions with us human users or even other AI agents.

Existing resources

We will leverage our in-house AI agent technologies and characters to demonstrate this benevolent AI interaction framework. This includes the AI agent models, processes, architecture, and how they run on P2P platforms (e.g., Telegram, or Discord).

See the Additional video section for a demo of one of our existing AI agent characters McDolan.

Open Source Licensing

We will open-source our Benevolent AI Agent framework using Apache License 2.0.

This license will not apply to our existing AI agents, including McDolan, which we have been developing independently.

Links and references

Temporai links

• Website link
• GitHub link
• YouTube link - currently videos are unlisted
• Telegram link - for you to test out our McDolan AI agent, only for human-agent interactions
• Discord link -  for you to test out our McDolan and other AI agents, can do both human-agent and agent-agent interactions

Existing projects links

• SYBIL open-sourced code link + example notebooks link
• Onboard NeuralProphet open-sourced code link + example notebooks link
• MAUQ example notebooks link
• Deep SYBIL example notebooks link 

Additional videos

Here is a YouTube video of me interacting with one of our existing AI agent characters McDolan, on the Telegram platform.

During the review stage, we will post more videos about the additional capabilities of our agents, including making/receiving crypto payments, multi-agent interaction, and more.

This shows that we have hands-on experience with agent interactions. This proposal can help us with making agent interactions more benevolent to humans.

Was there any event, initiative or publication that motivated you to register/submit this proposal?

Describe the particulars.

We are attending the ETHDenver conference, where there is a rapidly growing number of crypto x AI agent side events. This inspired us to build, improve, and showcase our own AI agents and systems.