The Complete Beginners Guide To Autonomous Agents
Everything you need to know.
Ok, let’s start with what you already know.
Artificial intelligence can be used to complete very specific tasks, such as recommending content, writing copy, answering questions, and even generating photographs indistinguishable from real life.
You tell the AI to complete the one task, it completes the one task. Simple.
But what if you don’t want to have to come up with all of the tasks for the AI to do? What if you want a teammate rather than just a tool? What if you want the AI to think for itself?
Like really think for itself.
Imagine you made an AI that you could give an objective to, even something as vague as “Create the best ice cream in the world”, and the AI would come up with a todo list, do the todos, add new todos based on it’s progress, and then continue this process until the objective was met.
This is exactly what “Autonomous Agents” do, and they are the fastest growing trend amongst AI developers, yet most people don’t know about them.
(At the time of writing this article, no major publications have written about autonomous agents, and since publishing, only a few have covered it, so if you’re reading this… you’re very early.)
What are autonomous agents? Why are they such a big opportunity? How do they work? What does this look like in the future? How can I build or use one? How can I meet other people interested in autonomous agents?
These are the questions I’m going to answer for you right now.
Ready? Let’s do this.
p.s. I am CEO and co-founder of Octane AI, where for seven years we have been building conversational AI products, and are more recently building generative AI and autonomous agent solutions for brands. In 2016 I predicted that around now chatbot interfaces would take off and start to replace standard website UI, and now over 100 million people use ChatGPT and websites like it. I am now similarly predicting that autonomous agents will be widely adopted in the future, but this prediction won’t take seven years to come true, it will happen blazingly fast.
p.p.s. After writing this article I showed the draft to 125 of the smartest and most interesting people I know, including Emad Mostaque (Founder of Stability AI), Tony Hu (Former Acting Head of Emerging Technology for the FBI, and founder of Bondoo AI), Troy Carter (Lady Gaga’s ex Manager), Sahil Lavingia (Founder of Gumroad), Elizabeth Yin (Co-Founder of Hustlefund VC), Hugh Howey (Author of Wool), Chris Yeh (Author of Blitzscaling), experts from NVIDIA, Meta, investors like Ryan Hoover (creator of Product Hunt) and Erica Brescia (Manager Director of Redpoint Ventures, prior Github COO), and many many more. Their thoughts and opinions are sprinkled throughout, they will give you unique insights shared with the world for the first time.
Want to watch a video instead of read an article?
What Are Autonomous Agents?
Autonomous agents are programs, powered by AI, that when given an objective are able to create tasks for themselves, complete tasks, create new tasks, reprioritize their task list, complete the new top task, and loop until their objective is reached.
Read that description above one more time, because while it is simple, it is wild.
Autonomous agents can be designed to do any number of things, from managing a social media account, investing in the market, to coming up with the best children’s book.
“And these are, like, real? These exist right now?”
Yes, I know it sounds like science fiction, but these are functioning and real. If you can code you can make one in just a few minutes. And it is only the beginning.
The programming techniques and the AI needed to power autonomous agents are real and extremely new. There are many open source projects, like AutoGPT, BabyAGI, and Microsoft’s Jarvis, that are trending on Github and within AI communities and departments.
In the first two weeks of the creation of open sourced autonomous agent code bases, almost 100,000 developers are building autonomous agents, improving them, and pushing them to their limits, and thats only in the first few weeks of these concepts being invented. The number of developers working with this technology is growing at an increasingly faster rate.
It has grown larger than long time popular codebases including laravel, bitcoin, django, and pytorch.
Auto-GPT Github popularity increasing exponentially, faster than any codebase in history
This is not science fiction. Many think these autonomous agents are the beginning of true Artificial General Intelligence, or commonly referred to as “AGI”, which is a term used to describe an AI that has gained sentience and become “alive”.
Check out this autonomous agent that was just released from HyperWrite, you can see it living in the browser and helping you order a pizza.
You just say “order a large plain pizza from Dominos to One Vanderbilt” and it just… does it.
HyperWrite’s autonomous agent controlling the browser to order pizza
Or, maybe even more impressive, check out this experiment done in collaboration between Stanford and Google where they created a virtual town of 25 autonomous agents, and told one of them to plan a Valentine’s day party.
The simulated people went about their days, talking to each other, forming new memories, and eventually most of them heard about, and showed up to, the Valentine’s day party.
From the research paper “Generative Agents: Interactive Simulacra of Human Behavior”
“Ok, uh, crazy… So autonomous agents are real... And you just tell it what it’s goal is and then after that it manages itself forever?”
You just give it the one objective, and the autonomous agent does the rest.
Just like a really good employee or teammate.
Although, if you wanted to, you could also design the autonomous agent to check in with you at certain key decision making moments so that you could momentarily collaborate on their work.
“But what can autonomous agents do, Matt? Like when you say they complete tasks, what the heck do you mean by that?”
In addition to analyzing their objective, and coming up with tasks, autonomous agents can have a range of abilities that can enable them to complete any digital task a human could, such as:
Access to browsing the internet and using apps
Long-term and short-term memory
Control of your computer
Access to a credit card or other form of payment
Access to large language models (LLMs) like GPT for analysis, summarization, opinion, and answers.
Also, these autonomous agents will come in all shapes and sizes. Some will operate behind the scenes where the user is unaware of what they are doing, while some will be visible, like in the example above, where the user can follow along with each “thought” the AI has.
“Matt, I’m reading what you’re writing, I think I know what you are saying, but can you write out an example in plain english so I can be sure I understand.”
Yes, of course.
(pssst… Are you liking this article so far? I’m thinking of making some YouTube videos too! Subscribe to my YouTube @MattPRD to get alerted when they are up.)
Here is a super simple example of how an autonomous agent could work.
Let’s say that there is an autonomous agent that helps with research, and we want a summary of the latest news about a certain topic, let’s say “News about Twitter”
We tell the agent “Your objective is to find out the recent news about Twitter and then send me a summary”.
So the agent looks at the objective, uses an AI like OpenAI’s GPT-4 which allows it to understand what it is reading, and it comes up with it’s first task. “Task: Search google for news related to Twitter”.
The agent then searches google for Twitter news, finds the top articles, and comes back with a list of links. The first task is complete.
Now the agent looks back at its main objective (to find out the recent news about Twitter and then send a summary) and at what it just completed (got a bunch of links of news about Twitter) and decides what its next tasks need to be.
It comes up with two new tasks. 1) Write a summary of the news. 2) Read the contents of the news links found via google.
Now the agent stops for a second before continuing, it needs to make sure that these tasks are in the right order. Should it really be writing the summary first? No, it determines that the top priority is to read the contents of the news links found via google.
The agent reads the content from the articles, and then once again comes back to the to do list. It thinks to add a new task to summarize the content but that task is already on the todo list so it doesn’t add it.
The agent checks the todo list, the only item left is to summarize the content it read, so it does that. It sends you the summary just like you asked.
Here is a diagram showing how this works.
From Yojei Nkajima’s BabyAGI
Pretty cool right?
And keep in mind that this is the very beginning of this new paradigm. It’s not perfect, it hasn’t taken over the world yet, but the concept is frighteningly powerful and with increased development and experimentation will quickly find it’s way into our daily lives.
So now you understand at a high level what an autonomous agent is, but why exactly are these such a big opportunity?
Let’s get into it.
Why Autonomous Agents Are Such A Big Opportunity
It’s pretty clear that soon you won’t only have the options of hiring humans as employees, you will have the ability to hire AIs in the form of autonomous agents.
And they are not going to be nearly as expensive as people are, they won’t sleep, they won’t quit, and they will work extremely efficiently.
These autonomous agents will exist in every industry and for every task imaginable.
These are just a handful of examples. Let your imagination run wild.
The list can go on and on. Anything a person could do, an autonomous agent will (eventually, but soon, and in some cases already) be able to do better.
So what do you do with this information?
There are two very real opportunities.
You create autonomous agents and make them available for others to hire.
You hire autonomous agents and can now afford to be more productive in your personal life, or in business.
Imagine a world where one person builds a company with only autonomous agents on their team. Within your lifetime you will likely see a one person team do this and reach a market cap of over a billion dollars, something it usually takes many many people working together to accomplish.
Right now in the early days there will be a period of time where early movers, either on making autonomous agents, or using them, will have a huge advantage against competition that is not yet leveraging these systems.
By reading this article you are already ahead of 99% of the world.
Let’s dive into more detail on how these autonomous agents work.
How Autonomous Agents Work
You’ve already read over a high level of how autonomous agents work, but I thought it would be helpful to give you one version of an overall framework, as well as break down a couple examples of autonomous agents step by step.
First, here a generalized framework for an autonomous agent:
Initialize Goal: Define the objective for the AI.
Task Creation: The AI checks its memory for the last X tasks completed (if any), and then uses it’s objective, and the context of it’s recently completed tasks, to generate a list of new tasks.
Task Execution: The AI executes the tasks autonomously.
Memory Storage: The task and executed results are stored in a vector database.
Feedback Collection: The AI collects feedback on the completed task, either in the form external data or internal dialogue from the AI. This feedback will be used to inform the next iteration of the Adaptive Process Loop.
New Task Generation: The AI generates new tasks based on the collected feedback and internal dialogue.
Task Prioritization: The AI reprioritizes the task list by reviewing it’s objective and looking at the last task completed.
Task Selection: The AI selects the top task from the prioritized list, and proceeds to execute them as described in step 3.
Iteration: The AI repeats steps 4 through 8 in a continuous loop, allowing the system to adapt and evolve based on new information, feedback, and changing requirements.
But, now lets apply it to a few different use cases I decided to extrapolate on.
Example #1: Social Media Manager Autonomous Agent
Let’s say that instead of hiring a social media manager to manage your social media accounts, instead you wanted an autonomous agent to do everything for you at a fraction of the cost and with round-the-clock intelligence.
Here’s what a framework for that autonomous agent might look like.
Initialize Goal: Set up the initial parameters, such as target audience, social media platforms, content categories, and posting frequency.
Data Collection: Collect data on past social media posts, user interactions, and platform-specific trends. This could include likes, shares, comments, and other engagement metrics.
Content Analysis: Analyze the collected data to identify patterns, popular topics, hashtags, and influencers relevant to your target audience. This step could involve natural language processing and machine learning techniques to understand the content and its context.
Content Creation: Based on the analysis, generate content ideas and create social media posts tailored to the platform and audience preferences. This could involve using AI-generated text, images, or videos, as well as incorporating user-generated content or curated content from other sources.
Scheduling: Determine the optimal time to post each piece of content based on platform-specific trends, audience activity, and desired frequency. Schedule the posts accordingly.
Performance Monitoring: Track the performance of each post in terms of engagement metrics, such as likes, shares, comments, and click-through rates. Gather user feedback, if possible, to further refine the understanding of audience preferences.
Iteration and Improvement: Analyze the performance data and user feedback to identify areas for improvement. Update the content strategy, creation, and scheduling processes to incorporate these insights. Iterate through steps 2–7 to continuously refine the social media management system and improve its effectiveness over time.
By incorporating this loop-type system in social media management, you can create a dynamic and adaptive strategy that evolves with your audience’s preferences and the constantly changing social media landscape. This will help to increase engagement, reach, and overall effectiveness of your social media efforts.
Example #2: Political Campaign Manager Autonomous Agent
What if you are running for political office and you want to leverage an intelligent and never-sleeping assistant to help you win?
This is what an autonomous agent that helps you win an election might look like.
Initialize Goal: Win the election by securing the majority of votes.
Data Collection: Gather data on voters, demographics, key issues, campaign messaging, and other relevant information.
Context Analysis: Analyze the collected data to identify trends, opportunities, and challenges. Refine the initial goal into specific subgoals based on this analysis, such as targeting undecided voters, increasing voter turnout in key areas, or improving campaign messaging on particular issues.
Task Generation: Generate tasks related to the refined subgoals, such as planning voter outreach events, creating targeted advertisements, or developing policy proposals.
Task Prioritization: Rank tasks based on their potential impact on achieving the subgoals and the overall goal of winning the election.
Task Execution: Execute the highest priority tasks, allocating resources and assigning team members as needed.
Performance Monitoring: Assess the effectiveness of completed tasks by tracking key performance indicators like voter engagement, public opinion, and fundraising metrics. Evaluate the success of individual tasks and overall campaign progress toward the subgoals and initial goal.
Iteration and Improvement: Analyze the performance data to identify areas for improvement. Update the campaign strategy to incorporate these insights. Iterate through steps 2–8 to continuously refine the political campaign management system and improve its effectiveness over time.
At first one candidate might use an autonomous agent and have a huge advantage over everyone, but then imagine what this looks like once every candidate has one… or many.
Example #3: Math Tutor Autonomous Agent
Here is an autonomous agent that is designed to teach a child math.
Initialize Goal: Identify the child’s current math skill level and set a personalized learning path to help them improve.
Data Collection: Gather information on the child’s learning style, progress, and performance through assessments, interactions, and feedback.
Context Analysis: Analyze the collected data to identify strengths, weaknesses, and learning preferences, as well as any external factors influencing the child’s progress.
Task Generation: Generate tutoring tasks based on the child’s needs and learning path, such as selecting appropriate exercises, providing explanations, or offering real-life examples and applications.
Task Prioritization: Rank tutoring tasks based on their potential impact on the child’s learning and skill development, ensuring a balance between challenge and engagement.
Task Execution: Execute the highest priority tasks, adapting the tutoring approach and content delivery as needed to maximize the child’s learning and engagement.
Performance Monitoring: Assess the effectiveness of the tutoring by tracking key performance indicators (KPIs) such as progress toward learning goals, improvement in math skills, and the child’s engagement and satisfaction.
Feedback Loop: Continuously monitor the child’s performance and update the context analysis, task generation, and task prioritization steps based on new data and insights. Adjust the initial goal and learning path as necessary to better support the child’s math skill development.
Iteration and Improvement: Analyze the child’s performance and update the context analysis, task generation, and task prioritization steps based on new data and insights. Adjust the initial goal and learning path as necessary to better support the child’s math skill development. Iterate through steps 2–9 to continuously refine the political campaign management system and improve its effectiveness over time.
This autonomous agent loop type system outlines a process for an educational math tutor to adaptively support and guide a child’s learning experience, focusing on continuous improvement and personalization based on the child’s needs and progress.
The Future Of Autonomous Agents
Right now humanity is in the very beginning of developing autonomous agents. We’re poking around, breaking things, experimenting, making bad things, making good things.
Barely any commercialized products have even been released, everyone is still in development mode.
But soon, that is going to change. Autonomous agents are going to start showing up all over the place until one day it will be incredibly strange for someone to not have one, or multiple, autonomous agents helping them out at any given time.
People will move through life with autonomous agents of all kinds augmenting their movements, decisions, and actions. If at some point we have neural implants then this will all happen seamlessly just like thinking in your own head works today.
Here are my predictions for the future of autonomous agents:
2023 multiple commercialized autonomous agents for gaming, personal use, marketing, and sales.
2024 commercialized autonomous agents for every category but not mainstream adoption.
2025 mainstream adoption of autonomous agents in every category for everything imaginable.
2026 most people in first-world countries are going about every day life with the support of an army of autonomous agents.
In the next 2-5 years most people will work for an autonomous agent instead of a human.
“This is a lot to take in Matt, the future is going to be wild. Where can I start with autonomous agents today though?”
This is the best question to ask. I have all the resources you need.
Let me show you.
How To Build And Use Autonomous Agents
You are now ready to jump headfirst into the world of autonomous agents. I’m going to give you the resources you need to get started building or using autonomous agents on your own.
I’m excited to see what you can do with this, and if you make something cool, I would love to check it out.
Building Autonomous Agents
You have a couple different options here.
Build It Yourself: Look at the framework I provided earlier and embark on a journey to build everything from scratch! You can definitely do this, it’s not a scary as it might sound. Some recommended software solutions are OpenAI’s GPT-4, Pinecone vector database, and LangChain’s framework.
Auto-GPT: This is a popular open source option created by Toran Richards. It includes options to connect to the internet, use apps, long-term and short-term memory, and more.
BabyAGI: Another popular open source option, this one created by Yohei Nakajima. While this one doesn’t connect to the internet yet, it is extremely elegant with under 200 lines of code.
Microsoft’s Jarvis: Very similar to Auto-GPT and BabyAGI, but much more robust and brought to you by Microsoft and HuggingFace.
Using Autonomous Agents
Ready to have your own agent? Here are some options.
Spin up any of the options in the build your own section above!
AgentGPT: Create and run an autonomous agent (AutoGPT) from a website, no login required.
HyperWrite Assistant: Add a chrome extension that lets you give your browser commands and the browser follows through.
Autonomous Agents & Agent Simulations (via Langchain)
No matter if you can code, or you don’t yet know how, I encourage you to take a few hours to experiment with these. It is not as complex or as difficult as it may seem and the quicker you get your hands dirty the faster you’re going to learn about autonomous agents.
The autonomous agent landscape is wide open for interpretation and innovation. 99% of use cases have not been created or attempted, the possibilities are endless and the opportunity is yours for the taking.
This space is moving incredibly fast, faster than anything I have ever seen before. Every hour it feels like there is new information, new experiments, and new releases.
So how do you keep up with it all?
I got you covered. Come with me.
How To Meet People Interested In Autonomous Agents
You are only at the beginning of your autonomous agents journey, and I know you are still burning with questions and ideas you want to share.
If you’re sitting there thinking any of the following then I have the perfect solutions for you:
“I wish I could stay up to date on new developments in autonomous agents”
“I have an idea for an autonomous agent, I want to share it with someone and see what they think!”
“I built an autonomous agent, I would love to share it with people!”
“I want to invest in people building autonomous agents”
If this sounds like you, and your autonomous agent curiosity has been sparked, here’s what you should do next.
Subscribe to my newsletter and subscribe to my new YouTube channel to continue to get more insights, news, and product thoughts on AI and autonomous agents. I have been building products (used by thousands of businesses) in this space since 2016 and try to always be on the forefront of what is happening.
Join the Autonomous Agents group on Facebook. Here you can share content, projects, and opinions on autonomous agents.
For example when I talked about autonomous agents to Emad Mostaque, the founder and CEO of Stability AI, his response was a coy “Swarm intelligence will beat AGI.” What does he mean by that? Subscribe to my newsletter and we’ll explore it deeper.
The world is changing fast and I am so excited to dive headfirst with you into merging humanity with artificial intelligence.
Build something people want. Try not to destroy the world on accident. I’ll talk to you soon.
p.s. Want to chat? I’d love to hear from you. Reach out on Twitter @MattPRD or send me an email at matt at mattprd dot com.
p.p.s. All the artwork in this post was made with AI! I have turned them into posters and they look amazing. If you want some for your wall you can check them out here.