Artificial intelligence (AI) innovation has been dominating the news space as well as our minds for quite some time, expanding the idea of what all is possible. After Stable Diffusion and the image generation trend came ChatGPT, which brought focus to LLMs (large language models) and this has only seen growth following the launch of GPT-4 and others. After LangChain brought attention to agents, autonomous AI agents emerged as the next obvious step.
In this article, we will dive deeper into the potential of autonomous AI agents and associated tools and resources.
What Are Autonomous AI Agents
Autonomous AI agents can be described as intelligent entities that can make decisions and conduct actions without human intervention in complex environments. Think of them as language model-powered bots that disintegrate problems and solve them iteratively to achieve specific goals, taking actions on behalf and in the interest of the users. These agents may utilize reinforcement learning which is a machine learning technique that enables them to learn from their experiences and adapt over a period of time. In reinforcement learning, agents use trial and error and receive feedback as rewards or penalties for actions.
In summary, an autonomous AI agent,
- Can be a simple program comprising only a few rules or can be a large and complex system
- Responds to the events and states in their environment
- Is independent of direct interference or instructions by the user/owner
- Works towards a given goal
- Is highly scalable
Multi-Agent Systems (MAS)
A multi-agent system is designed using several interacting agents and is more complex than a single agent. These systems can solve problems that may be difficult or impossible for a unitary agent. A MAS can not always be the same as the agent-based model (ABM). Some literature suggests that ABM is more frequently used in sciences whereas MAS finds applications in engineering and technology-related problems.
LLM Vs. Autonomous AI Agents
Consider this.
If you are making a reservation in a restaurant, then LLM can search for the best restaurants in a city of your choice. You can ask an agent to search for the highest-rated restaurant that has availability and book the table for you. An autonomous AI agent, on the other hand, can find the best restaurant in a given city that aligns with your preferences and schedule and book it for you and your friends.
These agents achieve this by dividing the task into different sub or small tasks and utilizing the memory from each step to guide the actions.
Artificial Agents Have Captured Imagination But…
Although in their current state, autonomous AI agents offer a myriad of possibilities, there is room for improvement until they are adopted on a large scale. Especially, when it comes to performance, quality of the output, and user control.
- Logical reasoning does not always translate to good execution
In theory, GPT-4 can handle chain-of-thought reasoning and break down a task into a process with multiple steps. However, when it comes to practical execution, agents may face challenges when executing their sub-tasks. Since there is not much external feedback, they find it difficult to take a step back. This means they may get stuck in executing the same task in a loop or imaging a step.
- High costs
Since the architecture of these applications utilizes recursive loops, there may be multiple repetitive calls of the LLM. Although the cost per call is low with tools such as OpenAI’s APIs, the in-house models may not be as economical.
- Learning curve
Autonomous AI agents are spun up and are usually not reused. This means they are not able to learn from previous attempts, prompts or their own errors. With that being said, services that help agents persist are in the pipeline which will make their management much simpler.
AI-Powered Autonomous Agents: Resources and Tools
Here are some important platforms you should know about.
Auto-GPT
Auto-GPT is an open-source Python application that relies on OpenAI’s GPT-4, which is a multimodal LLM. It chains together LLM “thoughts”, to autonomously achieve the set goals. It is autonomous and as opposed to ChatGPT requires minimal human interaction. It is, in fact, able to self-prompt.
It was developed by Significant Gravitas and you can access its open-source GitHub repository here.
You can watch the demo video here.
BabyAGI
BabyAGI is a pared-down version of the Task-Driven Autonomous Agent shared on Twitter on 28 March 2023. In a very short span of time, BabyAGI has inspired multiple projects that you can access here. It uses Open AI and vector databases including Chroma and Weaviate to create, prioritize, and execute tasks.
It was developed by Yohei Nakajima and You can access the GitHub repository here.
Image source: GitHub
JARVIS/HuggingGPT
It is a collaborative system comprising an LLM as the controller and different expert models as collaborative executors (from HuggingFace Hub). It consists of four stages.
- Task Planning: ChatGPT analyzes the requests of users to understand their intention, and disassembles them into possible solvable tasks.
- Model Selection: ChatGPT selects expert models hosted on Hugging Face based on their descriptions to solve planned tasks.
- Task Execution: Invokes and executes each selected model, and returns the results to ChatGPT.
- Response Generation: Uses ChatGPT to integrate the prediction of all models, and generate responses.
You can access the original research paper here and the GitHub repository here.
Image source: GitHub
OpenAGI
It is an open-source AGI research platform that has been designed to offer complex, multi-step tasks and is accompanied by task-specific datasets, evaluation metrics, and a diverse range of extensible models.
You can access the GitHub repository here and an introductory video here.
Image source: GitHub
Camel
It is a novel communicative agent framework called role-playing. It involves using inception prompting to guide chat agents toward task completion while maintaining consistency with human intentions.
You can access the GitHub repository here and the demo here.
Image source: GitHub
AgentGPT
AgentGPT enables users to create and deploy customizable autonomous AI agents directly on the web. It works by chaining language models (Agents) to execute a given goal.
It is being developed by Asim Shreshtha and you can access the platform here.
Image Source: AgentGPT
BabyAGI UI
BabyAGI UI has been designed to make it easy to run and develop with BabyAGI in a web app such as ChatGPT. This is a port of BabAGI with Langchain.js.
You can access the repository here.
Image source: GitHub
AutoGPT GUI
This is a graphical user interface to AutoGPT. You can access the GitHub repository here.
Image source: GitHub
Westworld Simulation
Google’s and Stanford’s researchers have created an interactive sandbox environment consisting of 25 generative AI agents that can simulate human behaviour.
You can read more about this project here.
Image source: Towards Data Science
Additionally, you can explore the following agents.
Here are some demos that can get your imagination running!
If you want to dive deeper into reinforcement learning, here are some resources you can use.
- Spinning Up in Deep RL
- Deep Reinforcement Learning through Policy Optimization
- Reinforcement Learning-An Introduction
- Stable Baselines- RL Made Easy
- Course on RL-David Silver
Final Words
Although autonomous AI agents are in the early stages, they offer immense potential and outcomes. We discussed some of the prominent projects and with that being said, these projects do have certain limitations and scope of improvement. Additionally, there are concerns related to autonomous AI agents such as the agents getting stuck in a loop or hallucinating a step apart from security and ethical issues. Nonetheless, these autonomous agents hold promise for the future!
