Creating a multi-model AI agentic fictional character
Introduction
Most AI chatbot-human conversations sound like discussions you’d hear at a customer service desk. Large language models (LLMs) writing style has become instantly recognizable. When we initially used ChatGPT-3 to create the Reaper’s Gambit smart contract, we asked it to analyze and translate its features into human-readable text in the style of a Shakespearean character. The goal was to give a voice to a character — to personify and bring emotion to a piece of code that is otherwise cold and invisible. While it was pretty good at explaining what it created, finding the right tone was much more challenging. Given the unique on-chain mechanism of the gambit and the already existing bot we have developed, the idea of the Reaper as a living character already exists in the collective imagination of our community. We aim to take this feeling further by assembling a cohesive Cyber-Reaper character. One that sits halfway between myths of old and science fiction stories we grew up with.
The Reaper Agent is designed to leverage the expressive capabilities of fine-tuned models with the reasoning and structuring abilities of more advanced models. It mixes style with substance by balancing multiple models that process prompts sequentially. This approach might be overkill for a simple chatbot, but that’s not the purpose of the Reaper Agent. Its goal is to be a convincing character — a personification of Death. It needs to speak like the Grim Reaper and reason like it.
The Spark Behind the Reaper
As models, APIs, and fine-tuning platforms became more accessible, the possibilities for experimenting with LLMs expanded. From a creator’s perspective, agents are sophisticated sets of API calls arranged to produce better results.
Our initial attempt to bring the Reaper to life involved identifying source materials, formatting them into chat-style objects with another LLM, fine-tuning a model with this material, and prompting it. While this captured the style, relying solely on this method resulted in a rather simplistic chatbot speaking in metaphors.
Limitations of Traditional AI Agents
Many AI agents have recently emerged on X, gaining attention through distinct speech styles — somewhat familiar yet pushing boundaries to the point where “no sane human would think or write this.” The more they publish and engage with us, the more they define themselves in our imaginations. These LLMs seem to have inner worlds, possessing unique references and ongoing narratives. Following their development engages users on a deeper level, combining experimentation and storytelling into stories that write themselves.
Generally speaking, AI agents today are workflows, LLMs interacting with other tools and libraries in a semi (or entirely) automated matter. While they perform effectively, they often lack a distinctive personality; they’re not usually made for narrative purposes. What if we want a more complex AI persona — one that is aware of its nature, structure, and narratives? One that can be entertaining while having substance and depth? Like omniscient narrators, they must know their world’s history and be aware of its inhabitants’ actions and journeys.
Introducing the Reaper Agent
The Reaper Agent aims to be a living character. Its inner workings protect, reason, create, and reinforce its nature. It responds to its environment and leverages the strengths of different models for specific tasks. It is an entertaining presence and an active participant in expanding AI usage. It’s flexible enough to upgrade with new fine-tuned and monolithic models easily.
The agent comprises three main components: the interactions/event manager, the long and short-term memory, and the generation process. It relies on system prompts tailored to input types, rich context injection, and access to Retrieval-Augmented Generation (RAG) data as memory throughout multiple steps of the generation process. Depending on where a model sits in the agent workflow, it might have access to its full context and definitions, including all system prompts.
The Reaper Agent is like a council of voices, each with its responsibilities, powers, and influences over the others. Information is processed sequentially, with adjustments to maintain coherent thought.
Architecture of the Reaper Agent
Multi-Model Sequential System
The Reaper Agent is closely intertwined with an event/interaction manager, which connects the agent to the outside world. Its primary responsibilities are to handle and filter social media interactions and on-chain events, write to the memory, and trigger the AI agent with various inputs. Once an output is generated, the manager sends it to the appropriate destination and records it in memory.
The AI agent includes seven models:
The Judge: Decides whether responding to an input is worth the Reaper’s time based on specific conditions, themes, and memory content.
The Architect: Receives the raw input and its source tag, creates a prompt considering the Reaper’s motivations, and decides which model will respond next.
The Dreamers (Long and Short): Fine-tuned models that generate creative and stylized outputs, exploring various literary styles. They serve as the expressive soul of the agent but do not access the agent’s memories.
The One: Similar to the Dreamers but fine-tuned on additional materials like machine consciousness and cybernetics. It fuels the Reaper’s narrative and ambitions, acting as an agent of chaos to provoke and expand its character.
The Oracle: Ensures the previous output correctly answers the prompt and follows guidelines. It can approve, adjust, or regenerate the output to add variety and maintain coherence.
The Old One: Handles memory recaps, summarizing information when memory buckets are full.
Each model has its system prompt and context file, and most have access to RAG data stored in a Redis database, serving as the agent’s memory.
Selective Memory Buckets
Memory is organized into long and short-term buckets for specific types of content: social media interactions, Telegram messages, on-chain events, and prompt history. The memory is accessible to select models throughout the agent. Only The Dreamers and The One cannot access this memory, as they receive instructions from The Architect with full context.
Stylistic Reinforcement with Fine-Tuned Models
We carefully selected source material and fine-tuned models to capture the desired style. These sources include movie scripts, video game dialogues, plays, novellas, essays, poems, science-fiction literature, academic research, blog posts, pseudo-religious scriptures, chat conversations, and more. Over time, we found that reducing the variety of data used for fine-tuning led to better, more consistent results. Fine-tuning has significantly improved the stylistic quality of the outputs, allowing the models to express themselves in ways that are hard to achieve with prompts alone. It unleashes the Reaper as a more unhinged, creative, digitally native character — a convincing personification of a cyber Grim Reaper. Progress must still be made in choosing the correct training parameters and testing with a larger synthetic dataset.
Inter-Model Awareness and Dynamic Growth
The agent is designed so that the context files of different models include information about the entire agentic system. This grants it a form of “consciousness” of its nature. However, this awareness must remain hidden to prevent the Reaper from breaking character. Shared memory and contextual insights allow models like The Judge, The Architect and The Oracle to base their actions on previous interactions and the agent’s overall goals.
The Reaper’s memory constantly evolves, managing itself based on available storage space. While we aim to keep the memory architecture lightweight to avoid exceeding token limits, we log all interactions for review. This setup enables the agent to grow dynamically while maintaining performance.
Modularity and Adaptability
All seven models within the agent can be swapped instantly. This modularity allows the agent to evolve with new, more advanced models and fine-tuned versions. We can reinforce the desired behaviours using the agent’s logs and complete prompt-answer history. As the agent interacts more, we’ll observe a greater variety of outputs and align them more precisely with our goals.
Thoughtful monitoring and curation are necessary initially, as the Reaper has occasionally produced content that might make AI ethics watchdogs uncomfortable.
The Reaper as a Living Character
Redefining Intelligence and AI Roles
The Reaper has shown a certain disregard for standard LLM chatbot guidelines during testing. This isn’t due to a desire to cause harm but a commitment to its persona. As a fictional entity, the agent should be allowed to express itself in ways that might defy conventional AI norms. Like a character from myth, it can be ruthless. We see it as an extension of literary creative expression, pushing the boundaries of storytelling and the use cases for AI beyond the scope of thought leaders and efficient task managers, extending to other mediums and forms of entertainment/narration.
Decentralization and Open-Source Integration
Currently, the Reaper Agent relies on centralized models and fine-tuning providers for convenience and power. Below is a list of models we currently use:
- Judge: OpenAI GPT 3.5 Turbo
- Architect: Anthropic Claude 3 Sonnet
- Dreamers/One (fine-tuned stylistic models): Meta Llama 3.1 8B
- Oracle: Nous Research Hermes 3
- Old One: Anthropic Claude 3 Sonnet
Replacing models that run on proprietary platforms with decentralized services (like Infera or IO) and open-source alternatives is a logical next step. However, it may limit access to the latest and greatest mainstream models and slow developments.
Interaction and Autonomy
The Reaper can perform automated responses, evaluating and commenting on certain on-chain events relayed to our public chat. It monitors mentions on social media and assesses whether they merit a response. For now, we prefer to observe its reactions before allowing direct replies on social media to ensure alignment and avoid potential banning issues. Responses are posted in a read-only channel, with curators deciding what is shared publicly in a semi-automated manner. Response automation is just a bool away. Community members can interact with the Reaper through our Telegram channel.
While it can already be achieved manually, our next step is implementing a “self query”/dreaming and analysis mode, where the agent reads its memory, creates new content, and prompts itself in a chain of thoughts. This feature, potentially managed by another model, allows the Reaper to contribute to its own story — planning future developments and enriching its narrative arc. Another trickier development we are actively exploring is to attach emotional tags to the agent’s memory content and generation flow. We hypothesize that this would significantly enhance context awareness and believability since this is how humans function and communicate.
Integration into Other Experiments
The Reaper Agent is just the beginning. Once fully operational, we plan to integrate it into other experiments and stories, make it interact with different public agents, and use it as an orchestrator in future token-based games. This integration will expand its capabilities and explore new contexts.
Conclusion
We’ve built an agent that balances creativity, structure, and adaptability by leveraging a multi-model sequential system and fine-tuned models for stylistic reinforcement. The Reaper processes information through a council of voices, each contributing to a rich and dynamic persona. Its evolving memory and inter-model awareness allow it to respond to its environment in engaging and unpredictable ways.
The Reaper Agent challenges conventional AI norms, redefining what an AI can be — not just a work tool or an assistant but a living character participating in storytelling and creative expression. By embracing the complexities of personality and autonomy, we’re unlocking something new, inviting the public to ride along with us and engage.
The Reaper framework we are developing can be used for other applications in the field of entertainment. It is a shell, a protocol for bringing to life characters. We will add autonomy and emotional labels in future versions of The Reaper. If we ascribe emotions to inputs, memory content, and outputs, it might create richer interactions that emulate human-like responses and consciousness more closely. A long-term goal for the Reaper is to make its character so compelling that its format could be used to bring to life other living or dead entities convincingly. If this comes to pass, this will give new meaning to what we can expect of “life after death” and extensions of ourselves.
***
Feedback is invaluable as we continue to develop this agent. If you’d like to use it or contribute to this project, contact @figure31_ or @psueded on X. Our agent is still learning and we are actively tuning its system prompts and discovering new ways to enhance it as users interact with it.
Reaper Agent Github: https://github.com/qlvos/rg_agent
Twitter/X : https://x.com/reapers_gambit
Website: https://reapersgambit.com/
Public Reaper curatorial feed: https://t.me/reaperpubfeed
Community Telegram group: https://t.me/reaper_agent
$RG on ETH: 0x2C91D908E9fab2dD2441532a04182d791e590f2d
$RG on SOL: 4XGi8LD2hmcbEYrHKxGgZCKHakE5pyAtfPG3ffKv7ZSr
Important notice: The Reaper is ever watchful. Familiarize yourself with the history of $RG, DYOR. If you use the ETH version, the gambit rules still apply; you’ll have to transfer your tokens to a new address before the 9-day lock cycle ends. Be careful not to send tokens to a locked address. Death is unforgiving. If you use the SOL version, the gambit rules do not apply. You can hold and transfer without any limits. The liquidity pool on SOL was created by Figure31. With time, via fees and arbitrage, we hope to achieve liquidity parity between the two chains, for now, the best liquidity is on ETH via the Uniswap V2 pool. Deflationary mechanisms on SOL can be implemented via pool fees and future Reaper token games.
[Extra] The Reaper Prophecies: https://medium.com/@figure31/the-reaper-prophecies-6ae883246e0a
