2025 could bring more agentic AI developments. Enterprises could embed agents in their workflows. It could also lead to an orchestration infrastructure and better reasoning models.
Listen to this article. This audio was generated by AI.
For many, a future with AI agents resembles the Marvel Cinematic Universe character J.A.R.V.I.S.
J.A.R.V.I.S., the acronym for Just A Rather Very Intelligent System, began as a natural language computer system created by Tony Stark, a fictional industrialist and brilliant investor. It later became an AI system that served as Stark's assistant. Much later, J.A.R.V.I.S. gained a synthetic body and became the android Vision.
While AI agents -- autonomous and semi-autonomous generative AI systems that can take action on their own -- might be far from having the ability to gain bodies, they might get close to or surpass J.A.R.V.I.S. at some point next year.
The growth in the popularity of AI agents in the latter months of 2024 mirrors how ChatGPT and other generative AI systems catapulted into and transformed the AI market in 2022. Vendors seemingly jumped from developing the latest large language models (LLMs) and AI chatbots to creating agents and action models.
Other vendors have also introduced AI agents to enterprises to automate various business processes. Analyst firm Forrester Research lists 400 vendors now building agents.
"There's a lot of excitement right now about them," said Craig Le Clair, an analyst at Forrester Research. "There's also a fair amount of risk when you're unleashing an automation that can proceed without human checks and balances towards the goal.”
The excitement and the risk mean that AI experts and vendors have many expectations for AI agents in 2025.
Eliminating confusion with real applications
One expectation is that while 2024 laid the groundwork and foundation, 2025 will be the year AI agents become enterprise-ready, according to AI market experts.
This means the confusion surrounding agents will go away, said AJ Sunder, co-founder and CIO of Responsive, a vendor of AI-driven proposal and response-to-proposal software.
"There's confusion between agents and automation, agents and RPA [robotic process automation]," Sunder said. "A lot of that confusion will go away. Then we'll start to see more agents deployed and being used in the real world."
Whereas RPA uses robots or bots to automate repetitive tasks without the use of AI, agents involve AI technology. RPA is deterministic and predictable, but agents are not.
"They are similar in that they're both digital co-workers," Le Clair said. "It's just that when you add AI to a digital co-worker, we call it an AI agent, more intelligent and able to understand context, kind of knows how not to get stuck."
Some real-world applications of agents will be in customer service; others will be in finance or fraud detection, Sunder said.
"Anything complex that requires AI memory, planning and executing multistep, complex tasks, I think agents are going to play a huge role in," Sunders said.
One complex application is video creation.
"A lot of these agentic AI solutions can be actually deployed in a way that they assist the video creation process," said Shahzaib Aslam, director of research at Colossyan, an AI video platform.
An AI agent can help put together an engaging video that provides a compelling argument and includes a call to action to get customers to do something, such as buy a product, Aslam said.
"This becomes a very powerful tool because it will assist you in putting together a video, which is higher engagement and has more success rate," he said.
Not only will agents play a role in different use cases and applications like video creation, but many will also begin to use them to address problems of scale, said Gartner analyst Tom Coshow.
Enterprises have two problems with scale. They either aren't doing a good job because they don't have enough people to handle it, or they have people doing a great job, and they wish they could have more volume of that great job.
Tom CoshowAnalyst, Gartner
"Enterprises have two problems with scale. They either aren't doing a good job because they don't have enough people to handle it, or they have people doing a great job, and they wish they could have more volume of that great job," Coshow said. "Those are two good use cases for AI agents. In 2025 we'll see people realize that they need to focus AI agents on problems of scale."
However, there are different levels for applying and using AI agents, said Peter van der Putten, director of the AI Lab and lead scientist at Pegasystems, a workflow automation and decisioning vendor.
At one end of the spectrum, AI agents can read, integrate and synthesize information, and come to a certain level of conclusion but not take any action. The other end of the spectrum is when the AI agent acts based on the information it synthesized, van der Putten said.
"The real success of agents is not about the intelligent capabilities of these agents themselves, but it's more how you embed them," he said.
However, he continued, most enterprises will have to try it out before they can see its value.
"I'm even more surprised sometimes what these systems are able to do," van der Putten said. "The only way to find out is through safe experimentation."
Better reasoning models
Another expectation with AI agents is that LLMs will continue to be their brains. This means LLMs will need to become better at reasoning so AI agents can perform their tasks better.
The idea is that instead of the model generating just one response to a query, it generates multiple responses and thinks through the steps to find the final response.
While this can be expensive because an enterprise is now running multiple inferences to create a chain of thought, it also makes the model reason better, Aslam said.
He added that this will be an area that the AI industry and academia will explore in 2025.
"This way of adding interpretability into the models would make a lot of sense, and we will see a lot more work and research go into this direction of scaling the compute at inference time and making the models arrive at their predictions in a systematic and reasoning manner, rather than just simply creating content," he continued.
Specific task agents
While it is likely that more agentic use cases and applications will appear in 2025, they won't eliminate the need for human intervention.
Nevertheless, with the advent of a new level of automation brought by AI agents, the fear that jobs will be eliminated remains.
Some in the industry say that while AI agents will have autonomy in 2025, it won't be full autonomy. In other words, AI agents will perform parts of an individual job but will not take over the whole job. For example, you might have an AI agent find the contact of a travel agency you want to use, but it won't be able to make the full reservation.
"We'll see agents not as independent, taking over full job profiles, but they'll take on portions of a person's responsibility or a portion of a process, and then work in conjunction with traditional automations, work in conjunction with humans and work in conjunction with other agents," said Mark Greene, senior vice president and general manager at UiPath.
Agents that take over one aspect of responsibility will be specialized and handle the task with a single-minded approach. This will make the AI agent more accurate in completing its task, Greene said.
"The narrower the responsibility, the more you can measure how effective it is," he said.
AI Agents infrastructure
Other than the rise of single-task AI agents, 2025 may also be the year of building the infrastructure for AI agents, said Olivier Blanchard, an analyst with Futurum Group.
Getting to the point at which AI agents are communicating with other agents or even performing tasks in concert with humans requires orchestration, Blanchard said.
"2025 isn't going to be the year when we see a fully developed agentic AI," he said. "2025 is sort of the year that we build the infrastructure for it. We build sort of the foundations for it."
He added that the vendors that might be instrumental in helping build the infrastructure are chipmakers like Qualcomm, Intel and AMD.
"Qualcomm's processors will mostly work with agentic AI on devices," Blanchard continued. Meanwhile, Nvidia's processors are currently geared toward working with agentic AI in the cloud.
"Nvidia's GPUs are already widely used to train AI models, which forms the basis of what will become the agentic AI layer," he said. "Two years from now, agentic AI will be a mix of cloud and on-device software interacting together."
Currently, Nvidia is mostly working with the cloud, while Qualcomm is on-device. On the other hand, device makers like Apple and Samsung will take part in creating the orchestration layer that allows agentic AI to work across platforms, devices, and individual apps, Blanchard said.
"We have ... the fundamentals of this," Blanchard said. "What we don't have is a sort of 'I can do everything' piece of it yet."
One way of reaching the orchestration layer is with multimodal AI. While generative AI systems like ChatGPT have the input-output function, they can't act on a human's behalf to connect to other apps.
However, as multimodal AI grows and becomes more mature to enable image input to lead to video output, then that will bleed over to agentic AI working better.
"As the models get smarter, that will make our agents smarter," Coshow said.
AI agents need an orchestration layer that works across different platforms and devices, Blanchard said. An orchestration layer is made up of the links that enable the AI agent to go from one platform or interface to the next, or from one application to the next.
If Qualcomm builds its orchestration layer and AMD builds its own, this will make interoperability for agentic AI a challenge.
"If all the chipmakers are using their own orchestration layer, they're not necessarily going to be able to talk to each other super well," Blanchard said.
Agentic Challenges in 2025
Similar to other AI technologies, AI agents will face challenges in 2025. One is data.
Because data is usually spread across different sources and processes, it might be challenging to give AI agents the data they need to perform the tasks they're being asked to do, Greene said.
Another problem is lack of knowledge about the design process for agentic automation, Greene added.
For instance, the industry will need to learn when humans should interact with agents, how they should interact, and what channels to use with agentic AI, he said.
There is also the challenge of trust, Sunders said.
"If the underlying technology still relies on generative AI and large language models, those shortcomings will be inherited by agents," he said.
Despite these obstacles, 2025 will be a big year for AI agents, he continued.
"We will figure out where agents make sense, how to deploy them, how to earn the trust, before we go completely hands up," he said. "This promise that it can be entirely autonomous, I do think will happen; whether that will happen in 2025, I don't think so."
Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.
