Using AI Agents Pt 2: AutoGPT and SuperAGI
AI Agents are better and more useful than ever, but still lack reliability
AutoGPT and the Rise of AI Agents
This article is a follow-up to “Using AI Agents pt 1: BabyAGI, AgentGPT and Beyond.” In this article we will discuss AutoGPT and SuperAGI.
The release of GPT-4 in March inspired tinkerers to push further in building useful applications, and in particular the development of autonomous AI agents based on GPT-4 created a lot of excitement and hype. One of the first agents to grab attention by leveraging GPT-4 was AutoGPT. Since then, a plethora of AI Agents have arisen and the area is rapidly evolving and improving.
AutoGPT is an advanced autonomous AI agent developed by Toran Bruce Richards. It is designed to generate prompts for the underlying language model, typically GPT-4, to execute tasks automatically without human intervention based on a predefined goal. The power of GPT-4 as an LLM makes for a simple yet powerful AI Agent concept:
Create a goal or series of goals. You might add constraints as well as access to tools, such as web browsers.
Call the LLM recursively to break down the goal into instructions, and iteratively execute those instructions on the LLM until complete.
Automatically calling the LLM iteratively chains the execution together to create a simple autonomous agent. That agent can continue to work and improve its result until it’s done.
AutoGPT has interesting and even amazing capabilities. Autonomous text-generation capabilities can be used for storytelling and creative writing. Users can string together larger bodies of text than could be done in a single prompt, such as plots, dialog, ad copy or other written works. AutoGPT can iteratively work on improving code. With web browser access, it can scrape the web for information, and then autonomously does the task for you.
The promise of AutoGPT is counterbalanced by its limitations, which applies to AI Agents generally. Overall, they are not yet reliable, and they can fall into infinite loops. They are limited by the quality of prompts and the responses; if the LLM used cannot reason well, write well, or code well, iterating on it doesn’t yield fruitful results.
SuperAGI
While AutoGPT was the first, it may not be the best, so let’s turn our attention to the tool called “AutoGPT on Steroids," SuperAGI.
SuperAGI is a ‘developer-first’ environment and bills itself as “Infrastructure to Build, Manage & Run Autonomous Agents.” When the first part of the instructions for setting it up are “Go install docker”, it’s clear you’ll need to be a developer or hacker to wade into this. It’s not quite ready for end user consumer. That said, its not too hard to get set up if you follow instructions.
The installation and configuration do require some time and effort, but once set up it is fairly quick to run. It can be run on a typical laptop, but you also need OpenAI tokens, which can cost money if you use this extensively.
SuperAGI has a rich environment and framework to create a variety of different AI Agents. While under active development, it already has enough features to be considered an effective application. These features include pre-built tools for writing code, producing emails, accessing Github, and a plugin for a slack bot. These tools can be given to individual agents, which can be customized to solve certain tasks.
This figure below shows the SuperAGI interface to available toolkits that can be connected to agents.
SuperAGI has settings that allow you to set the number of iterations and time between steps. This feature prevents falling into infinite loops, while also giving the ability to fine tune how long you want your agent to function.
You can customize your database and long-term memory in SuperAGI to fine tune your agent, turning long-term memory (vector database) on or off. The default vector database for long-term memory is Pinecone.
One interesting and helpful thing about the SuperAGI interface is that it displays the agent’s thought process as it works. The agent will generate its output (‘thoughts’), which are broken out into ‘text’, ‘reasoning’, ‘plan’, ‘criticism’, and ‘speak’. With this output, the user can follow the agent’s line of reasoning and its plan for the given task. The agent will also give itself criticisms, such as gaps in its plan or need for more information.
SuperAGI, especially with its toolkits, has several useful applications. It can assist with or even automate entire processes such as writing code or emails, and its open source code can be picked apart to be better understood.
Email agents necessitate heavy interaction with the configuration of the toolkit, requiring email address and password, SMTP ports, and IMAP servers among other things. Once configured, it has the capability to read through emails, as well as write them and attach files.
The configuration set up can be fragile. When I gave it a task of writing an introductory email draft, it tried to read emails that didn’t exist. This can likely be attributed to unclear goals, and with more iterations and tighter guidelines it should be able to produce a cohesive email in my draft inbox.
Slack agents require a Slack bot token, but once configured can write messages to the chat its plugged into. The bot can be directed to answer certain questions, include quips, and basically anything else the agent’s underlying LLM can do.
As with other AI agents, SuperAGI isn’t always the most reliable. It can often be finicky based on instructions, or even misinterpret clear guidelines. For one example, when I removed the ability to read emails and specified not to read emails, the agent still made attempts to read emails because it believed it was lacking context. Somehow, it didn’t connect or follow its full instructions or constraints.
SuperAGI shows promise as an AI application and can do some useful tasks. However, in our use so far, we’ve found it not very reliable, so it’s more of a fun AI project and not ready for prime-time consumer or business use. SuperAGI is also still in development and continues to improve, so the AI Agents will only get better from here.
Our Future with AI Agents
Looking ahead to when the reliability kinks are worked out, more tools and plugins are brought into the AI agents, and the underlying AI models are improved, we can expect AI agents to address an ever-growing set of real tasks in the workplace and in our daily lives.
The rise of AI agents will require workers to adapt to the new environment, learn new skills, and be more efficient leveraging these AI agents. The use of AI will vastly improve productivity and efficiency, and autonomous agents will be an essential part of making it happen.