While chatbots like ChatGPT, Claude, Gemini, etc. have dominated the AI space recently, these are not the endgame for most companies. Most companies want to produce AGI (Artificial General Intelligence) – artificial intelligence that can reason as well as, maybe even better than, humans. But there are a few stops along the way.

Chatbots have shown tremendous potential, but there's a cap on how useful they can ultimately be. Without autonomy, there's only so much efficiency and productivity chatbots can bring to the table. And how can that pass in a world that loves efficiency? It's the reason chatbots aren't bringing in the expected numbers in revenue. But chatbots are the very bottom of rung of the ladder.

That's why AI companies are placing their bets on AI agents as the most immediate future of AI. AI agents won't be like chatbots, nor like the bots you see these days in the form of customer support agents on most websites. They will be able to go beyond following a set of instructions and make decisions.

Interacting with these customer support bots, in their current state, is an exhausting experience at best. They can mostly never solve your problem and if they can, it's a long process; a human agent on the other hand is always more efficient and capable.

But with autonomous AI agents, that'll change.

What are AI agents?

Even the experts don't have a definitive definition for it. What they have is a vision that's evolving continuously.

But some things are clear. AI agents are models that can autonomously make complex decisions in a real-world setting. They'll still need occasional human guidance and intervention but the breadth of tasks they'll be able to undertake would be massively greater than the current chatbots.

Chatbots like ChatGPT can help increase human productivity. However, AI agents would be able to replace humans, at least at the lower levels.

Unlike current generative AI bots that work only by generating the next likely word in a sentence, AI agents will be able to think and reason. We're already seeing a preview of a reasoning model from OpenAI in the form of o1.

But the ability to think and reason is just one aspect for AI agents.

One of the characteristics that AI agents are supposed to have is the ability to pursue goals set by humans without being instructed, preferably in a complex and dynamic environment. So, unlike current chatbots where humans have to dictate every step, with AI agents, you'd only have to set the goal.

It's also important for an AI agent to be proactive and not wait to be prompted like a chatbot.

Another important aspect is that AI agents would also be able to learn from feedback, constantly improving and getting better at their jobs. However, they wouldn't need human reinforcement for feedback and getting better; they can learn from their own experience.

How does an AI agent work? Consider this example. When they encounter a problem, such as a customer complaint, they don't just follow a generic line of instructions with a pre-determine outcome. In such an example, they are capable of looking up a customer's reference ID and internal documents, asking further questions if required to understand the problem, and then providing a solution. They can even potentially hit up a human supervisor when they're stuck or if their programming requires them to get approval for certain actions. Finally, if they can't help, they can refer the customer to a human agent.

Applications

Customer support isn't all AI agents would be good at but it's definitely one of the first fields they're going to become prevalent in. Software development is another field that's going to be dominated by AI agents soon.

According to predictions, merely over the course of the next three years, many organizations would be using AI agents to write and rewrite code, where developers would be moved to the role of reviewers.

Many companies are developing agentic systems for their internal workload. They are already going from Proof of Concept stage to Pilot stage. Agents are nothing new; businesses have often used them to automate tasks. But with the involvement of AI, the tasks agents can handle will be more diverse and less specialized.

AI agents are already being deployed by many companies internally for various tasks and some are already offering initial versions to business customers, like Agentforce from Salesforce. It's only a matter of a few years before AI agents are slated to replace the call center workforce completely.

Most companies would also implement a multi-agentic system, where different agents are deployed to handle different tasks. These agents would be able to communicate and collaborate with each other.

While businesses would find themselves increasingly adopting AI agents in different roles, they won't just be limited to businesses, though. We're already seeing companies trying to improve personal assistants to become more useful with AI.

An ideal AI agent would be like a human assistant; it would be able to make purchases on your behalf, book and manage your travel, or schedule meetings and send invites. It should be able to use and interact with other tools, like web search, programming, or even other AI tools, in addition to human elements.

Ideally, AI agents will also be multimodal, like Project Astra demoed by Google at their annual I/O conference this year. It will be able to natively process audio, images, and video as input.

There won't be a single type of AI agent, though. Different situations would demand different sets of skills.

Current limitations of AI Agents

Currently, there are many problems on the path to getting fully autonomous agents.

For an agent to be truly autonomous and helpful, it'll need to make way less mistakes. Currently, AI is extremely prone to hallucinations. That number would need to come down to at least less than 1% for wider adoption. Getting the number down to less than 10% might be easy, but it's nailing the last digits that'll prove to be the most challenging.

Also, consider the previous example of an AI agent handling customer support. The final problem is one of the serious problems to tackle before the full-fledged implementation of AI agents. Ideally, they'd know when to pass off the problem to a human agent instead of just keep trying to solve it themselves, increasing the cost incurred. In short, their perseverance needs to be addressed when they're reasoning themselves instead of following a set of instructions.

Another problem AI agents face is that of context. Consider using an AI chatbot for coding in it's current form. Currently, it's easy to get underwhelmed by AI's coding abilities. They cannot produce long-form code, mainly because of constraints around context.

There are also security concerns and access controls to keep in mind before AI agents can become a full-fledged reality. With increased autonomy, these risks become even bigger. Businesses need to make sure that AI agents are only executing actions they are authorized to execute. They should also not be able to access any information they aren't allowed to.

Then, there's security vulnerabilities like prompt injection to be careful against.

The amount of training data and computing resources required are also a barrier, but according to a somewhat cryptic answer from Sam Altman, they might already have an answer to the problem of the training data.

AI companies are working relentlessly on realizing an agentic future and most of these problems would be addressed soon. For instance, Google is already providing a 2M context window, and already working on making it infinite.

That's to say that while AI is not as competent as we'd like it to be (or not like it to be) currently, the day isn't too far off. In fact, it's closer than most of us think.

If you've been thinking that there's a long time before AI will become actually useful and capable of handling the tasks you do, think again. AI agents are going to be here rather soon. It's high time to upskill yourself and adapt for roles where AI agents will likely be your co-workers. While there's still time before AI agents can autonomously handle great responsibilities, most companies are looking to start using AI agents at some level from next year itself. With their autonomous decision making, proactiveness, adaptibility and the ability to handle complex environments and goals, gear up for an AI agentic future.