Today’s AI agents are capable of perceiving, making decisions, and taking actions independently. With the rise of large language model (LLM)-driven AI agents, we are on the brink of a new era: AI agents may form their own societies and coexist harmoniously with humans.
Below, we will discuss what AI agents are, why they are emerging now, and how they will develop. We make 12 significant predictions about the impact of AI agents on businesses and consumers in 2024, hoping to help people gain an edge in the upcoming Agent technology revolution.
Image source: Wang Knowledge
Author: Zhang Changwang, Wang Knowledge
1 – The Dawn of AI Agents
2023 has been a year for generative AI (GenAI) and large language models (LLM), which generate content in unprecedented ways. In fact, there has been an explosive growth in the market for new models and applications. However, LLMs are just the beginning of the GenAI revolution.
LLMs are impressive in generating text, images, and code outputs that people can use in their work (or just for entertainment). But what if we could go further, using LLMs not just as the final step in generation but as the glue to complete more complex tasks? For example, if we could enable LLMs to engage in work that requires continuous steps, specialized tools, up-to-date information, and specific skills, this is where AI agents are born.
Composition of AI Agents. Image source: https://arxiv.org/pdf/2309.07864.pdf
AI agents are programs that can pair LLM models with various data and tools to accomplish tasks and get work done. Agents go beyond content generation to decision-making and taking action. Simple AI agents can interpret high-level instructions, formulate plans, and then execute steps to achieve goals. At each step, the AI agent can assess its progress and adjust its strategy. Crucially, these agents can access a set of supporting tools and data that they can choose to use.
AI agents will have ways to complete actions. Imagine an email agent that can not only generate email text but also automatically converse with recipients to achieve goals. Or envision a scheduling agent that can converse with multiple parties to gather availability and interest information, then schedule (or reschedule) events based on feedback. Combining intelligent decision-making with real-world actions in the form of AI agents will yield powerful and valuable outcomes.
2 – Building AI Agents
In some ways, the popular OpenAI chatbot ChatGPT is a type of AI agent, aimed at providing valuable conversations and actions in response to prompt inputs while using recent memory as context. Recently, chat tools like ChatGPT have also been integrated with tools such as web search and databases to leverage the latest information to enhance queries.
The fact is, the building blocks for future AI agents already exist. First, thanks to 2023, we now have a variety of excellent LLMs and coding tools (langchain, HuggingFace, PromptAppGPT, etc.) that work with them. Second, many existing software programs and functionalities are connected to and operate in the real world.Third, valuable data is everywhere. We just lack the experience to integrate these parts (LLMs, tools, and data) together, but this will develop quickly. As businesses understand their value and see the impact around them, the demand for building AI agent technology will grow rapidly.
Image source: Wang Knowledge
While these parts are already in place, the technical design of AI agents still requires a lot of thought.Developers need to decide which program tools the agents can access and how the agents will interact with them. They also need to determine what data is available and how it will be presented. For example, is it JSON, SQL, or a vector database? Is there accompanying metadata? Is the data intelligently chunked? How is the data updated? And so on.
Even after figuring out the tools and data, there are still many technical decisions to be made.What prompts will AI agents use, which LLMs can the agents call? What will be done with the outputs? How will agent performance be evaluated to make objective improvement decisions?
On the business side, assembly is also needed. Like others, product teams are also novices when it comes to GenAI and AI agents. Managers need to gain experience in building teams with the right expertise that combines domain understanding, product experience, software development, and AI skills. Before achieving balance in production applications, companies may also need to conduct proof of concept trials.
Many companies already have this vision. Leaders are preparing to act, and developers are gaining experience. Over time, more specific tools and business experiences will be developed. Users will gain confidence when interacting with AI agents, driving a faster technology iteration improvement cycle.
3 – Outlook on AI Agent Trends in 2024
Image source: Wang Knowledge
We expect 2024 to be another significant year for AI, as many visionary companies will learn and implement AI agent technology. Here are some of our predictions for AI agents in 2024:
1. AI Agents Performing Routine Tasks:AI agents will soon graduate from novelty toys to performing simple, mundane routine tasks.Tasks such as updating documents, scheduling, and auditing will be handled by new AI agents. These are accessible wins for businesses looking to venture into AI agents. The victories here may seem trivial, but they will represent a significant step from the hype surrounding AI to concrete evidence of business involvement.
AI agents performing routine tasks. Image source: Wang Knowledge
2. Specialization of AI Agents:The foundational generative models built and trained by large tech companies are general-purpose, and the out-of-the-box tools they use (like web search) are also typically general-purpose. By 2024, I expect to see more specialized models, as well as specialized tools and datasets for AI agents. I also anticipate some modularization here, especially in the case of complex tools and proprietary data. Some of these may be offered as services via APIs or provided in datasets.
Specialization of AI agents. Image source: Wang Knowledge
3. Multi-Agent Development:Do not expect all AI agents to be lonely workers in isolation. A natural step after specialization is to have specialized agents work together to accomplish more complex tasks than any one could achieve alone. Furthermore, multi-agent frameworks will begin to utilize hierarchies, where some agents focus on high-level goals while others handle task-specific work and report upwards.
Multi-agent hierarchical development. Image source: Wang Knowledge
4. AI Agents Consuming Data:In the future, many more page views may be completed by agents. Websites or APIs will need to be optimized for AI agents to discover and utilize. Currently, people typically think about the web and many other software tools from a human consumption perspective, but as AI agents consume more content and use more tools, this will change.
AI agents consuming data. Image source: Wang Knowledge
5. More Trust and Power to Choose and Act:AI agents will evolve from simple choices to more impactful decisions. As businesses progress, they will learn to trust their AI agents. These agents may be given budgets and funds to spend. Perhaps they will take on tasks such as purchasing stocks or launching personalized ads, as well as time-sensitive interactive transactions. For example, imagine negotiating with an AI advertising agent that can make decisions.
AI agents making impactful decisions. Image source: Wang Knowledge
Advertising agent: “Hey, Dan, if you buy our new product now, you can get a 20% discount.”
Me: “If you can ship it within three days, I will do so, and you can also send me some of your product stickers for free.”
Advertising agent: “Agreed! I will also send 3 gadget stickers. The shipping cost is x dollars, and we will ship to your address. Confirm to start payment.”
6. AI Agents as a Service:For years, Amazon Web Services has offered a service called Mechanical Turk, where businesses can hire large numbers of (human) workers to perform repetitive daily technical tasks, such as data labeling. Imagine being able to rent AI agents by the hour to accomplish similar tasks at scale on demand. These agents could even become teachable or adjustable so that they can perform better in selected fields or tasks. Are you ready to have AI agents review your contest submissions? Or, could you consult an on-demand financial advisor who can quickly adjust your funds to meet your changing goals?
AI agents as a service. Image source: Wang Knowledge
7. Some Jobs Will Face Pressure:There is no doubt that AI agents will continue the trend that GenAI has already begun, taking some jobs from some people. Consider the following reference dimensions:
-
Your job is done on a computer.
-
Your job is routine.
-
You can easily swap with others who have the same skills.
The higher your score on these dimensions, the more likely your job may start to feel pressure from AI agents. Many white-collar jobs are on the list — roles such as auditors, technical writers, and customer service agents can be automated to some extent. However, expect AI agents to enhance rather than completely replace many jobs in 2024.
Agents will put some jobs under pressure. Image source: Wang Knowledge
On the other hand, consider jobs that require the following dimensions:
-
Important tasks away from computers or digital interfaces
-
Diverse tasks
-
Highly specific knowledge or skills that few possess
In these cases, your job is relatively less affected by AI agent automation.
8. New Job Roles Will Emerge:As companies begin to use AI agents, they will need employees who can care for and interact with them. Will the next hot position be agent engineer? Agent manager? Agent trainer? Agent evaluator? Agent auditor? Perhaps AgentOps? It is hard to say exactly what the titles will be, but AI agents will need human counterparts who know how to interact with them, implement them, maintain them, and improve them. AI agents will also need monitoring and oversight.
Agent engineer? Agent manager? Agent trainer? Image source: Wang Knowledge
9. Hype and Doom.If there is one thing you can count on not changing in 2024, it is the crazy claims about AI. Those eager for attention (and advertising dollars) will continue to stoke the flames of our excitement. We will learn how AI agents will solve all our problems while also bringing about the end of humanity. If it sounds too good (or bad) to be true, it probably is. Unfortunately, this will make it difficult (just like in 2023) to discern where the true capabilities of agents lie.
Speakers and writers will continue to hype AI. Image source: Wang Knowledge
10. AI Governance.This has also been a feature of 2023, and calls for AI governance will continue into 2024. The power of AI agent technology will become more apparent, and the need to organize how we use it will also become more evident. However, governments will continue to struggle to guide the development of AI. Their leadership will be limited and easily influenced by lobbying groups seeking to win in the competition. Any entity that acts as if it controls GenAI is either delusional or wishful thinking. As mentioned above, the building blocks for AI agents are already freely available. Many tech leaders do not understand AI, so we may deviate from rationality in governance. Unfortunately, 2024 may not be a year of effective AI governance.
Governments will continue to struggle to guide AI development. Image source: Wang Knowledge
11. Intense Competition.In 2023, major cloud providers and AI companies engaged in fierce competition in the GenAI space. This situation is expected to continue like wildfire, with more AI companies releasing better products in 2024.Look for some useful AI agents from major tech companies. We will also see competition benefiting companies as implementers and consumers, as they turn the work of AI agents into business value that can be passed on to downstream customers.
Major cloud providers and AI companies are fiercely competing in the GenAI space. Image source: Wang Knowledge
12. AI Agents Supporting Multiple Devices:Due to AI agent technology, some physical devices are expected to become more powerful and interactive. Smart devices will move towards Smart 2.0.For example, I currently have a smart vacuum cleaner that can clean some rooms in my house. I do not talk to it now because it is not that smart. Besides reporting errors, it does not really communicate with me. With AI agent technology, I might be able to tell a Smart 2.0 vacuum cleaner, “Hey vacuum cleaner, can you clean this room in an hour? I’m busy right now.” Or “Can you pile all the toys over there and help me clean them up, then vacuum them?” I’m just brainstorming here. Perhaps the first widely used AI agent-supported physical devices won’t be vacuum cleaners, but I am very sure they are on the way. I look forward to seeing them, using them, and understanding their capabilities and limitations.
AI agents supporting multiple devices. Image source: Wang Knowledge
4 – Long-Term Trends for AI Agents
Image source: Wang Knowledge
Let’s broaden our thinking a bit; there are some longer-term trends that AI agents may develop. The following ideas may not appear in 2024, but it will be interesting to see if they do.
1. AI Agent Malware:Computer viruses evolve in various ways, so it is only natural that we will see a virus driven by AI agents. Additionally, AI agents are also involved in identity theft. We may need better methods to detect them, as they can be more persuasive than many human criminals. It is important to clarify that I am not predicting that these agents will generate themselves, but rather that human actors will create them.
AI agent malware. Image source: Wang Knowledge
2. Self-Destructing AI Agents:Honestly, in certain applications, on-demand APIs do not make sense, and AI agents need to be downloaded to local devices. Once a task is successfully completed, the agent can self-destruct to free up space. Another application might be to execute licensing agreements and allow customers to make regular payments.
Self-destructing AI agents. Image source: Wang Knowledge
3. Self-Improving AI Agents:There are various ways to improve LLM responses—few-shot prompting, fine-tuning, RAG, RLHF—but these improvements typically require manual implementation of components. Self-improving AI agents will recognize when they have learned something worth internalizing; they will be able to automatically adjust their LLM models to enhance their native performance.
Self-improving AI agents. Image source: Wang Knowledge
4. AI Agent Scientists:Compared to AI agents with business objectives, AI agent scientists may have learning tasks and access to tools and data. AI agents will formulate hypotheses and conduct tests. Perhaps agents can even request access to new tools. Connecting to real laboratories (e.g., materials science or biotechnology) could enhance experimentation and R&D.
AI agent scientists. Image source: Wang Knowledge
5 – Conclusion and Outlook
LLM applications stole the spotlight in 2023. We expect AI agents to shine in 2024. Right now, many agents (and LLMs) are like toys—fun to play with but not necessarily useful in the real world. However, the capability to build agents already exists, and we are exploring how to combine them. Companies that manufacture AI agents will accelerate their maturity and then release complex production-grade AI agents capable of working in real-life scenarios. Whether you like it or not, AI agents will become a part of our experience.
AI agents will become a part of our experience. Image source: Wang Knowledge
Thinking about the future with AI agents can evoke both excitement and fear. The future is coming too fast to keep up. The ability of AI agents to perform tasks and complete work presents tremendous opportunities for many individuals and businesses. We will be able to do more, create more, and enjoy more with less effort. However, these innovative new ideas will undoubtedly disrupt the businesses we know, and we all need to learn and adapt to new technologies.
References:
-
Comprehensive Survey on AI Agent Development and Applications in 2023: Concepts, Principles, Development, Applications, Challenges, and Outlook
-
Comprehensive Interpretation of OpenAI AI Agents: Principles, Planning, Memory, Tools, Cases, Challenges
-
Ten Thousand Word Review: Progress and Potential of LLM-Driven Agents by Fudan + Mihayou
-
https://medium.com/eviden-data-science-community/the-ai-agents-are-coming-a85b8fb1702d