More reasoning, interoperability key to future of agentic AI

The future of agentic AI

Eventually, agents will be ubiquitous. They will be built into work applications, connected to one another to interact, and as transformative as the internet was in the 1990s and smartphones were in the early 21st century, according to Dwarak Rajagopal, vice president of AI engineering at Snowflake.

Both the internet and smartphones made access to and use of information more widespread in their own way, and so will agentic AI.

"The internet and mobile devices democratized access to data, and agents will democratize actions on top of the data," Rajagopal said. "Everyone will be empowered to do things without having to think about where their data silos are. All of that will be solved by agents."

In effect, agents will serve as digital teammates that will work in concert with humans, according to Leskovec. They will enable people to be more productive while letting humans focus on what they are good at. For example, agents will take on tasks such as supply chain optimization that are in constant flux and can benefit from a level of oversight that would occupy too much of any person's time to allow them to do other work.

"What is going to happen is the notion of continuous decision-making," Leskovec said. "They will continually optimize, getting information and streamlining processes in real time."

Like Rajagopal, Leskovec noted that in the future, agentic AI will enable more people to do things that only experts have been able to do until now, while simultaneously enabling experts to do things that now take months in a much shorter time.

"There are things that take super advanced scientists months of work, and with agents, it will become much easier," he said.

Sangani also highlighted the potential for agents to augment what people can do. There is widespread fear that AI will take jobs away from people, he noted. But rather than take jobs away, AI will make people better at their work by taking on menial aspects of a given job -- for example, in computer science, assisting with developing and analyzing code when building an application.

"AI will amplify human potential," Sangani said.

Beyond the workplace, agentic AI will make access to services such as healthcare and education more even across geographic and economic boundaries, he continued. Students in rural areas will have the same access to knowledge as someone at an urban private school. Similarly, people will have access to elite healthcare no matter where they live.

"It is going to democratize access to everything across the world, and all the major companies should work toward that," Sangani said.

Improving models

While agentic AI represents the future of work, it's already part of the present.

Given data's role as the training material for AI, many data management vendors are developing environments for customers to connect their proprietary data with generative AI (GenAI) models to develop AI applications, including agents.

Snowflake is among them, having unveiled Cortex Agents in preview in February. In addition, Databricks and Microsoft are among those enabling customers to build agentic AI applications. Meanwhile, many vendors are also adding agentic AI capabilities to their own platforms to assist with tasks such as accessing and modeling the data needed to build agents, monitoring data to ensure its quality, and surfacing insights that humans might not otherwise discover.

But just as accuracy was a concern when interest surged in GenAI development following OpenAI's November 2022 launch of ChatGPT, which marked a significant improvement in the technology, accuracy is a concern with agents capable of acting without human participation. For agentic AI to become more widely used, the underlying models need to improve, according to Sangani.

Models such as Meta's Llama family and those that power OpenAI's ChatGPT and Anthropic's Claude are the foundation for agentic platforms, providing the reasoning that separates agentic AI from other forms of AI.

"If your model is able to understand and interpret really well and then do the reasoning, then your agentic platform will work really well," Sangani said.

Additionally, making sure that models have strong tool-calling capabilities -- the ability to interact with external systems to acquire knowledge beyond their built-in training -- to perform tasks is key to the future of agentic AI, he continued.

"When there is a single-tool chat, [models] work fine," Sangani said. "But when there is a multi-tool chat, which is what agents are supposed to do, that's when most of the models are failing. If models can supply the right kind of attributes, that will make agentic platforms stronger."

Importance of interoperability

Beyond improving the models that provide the foundation for agentic AI, the future of agents depends on their ability to work together across an enterprise -- just as someone in marketing might consult with someone in sales to come up with a plan for a campaign -- without human intervention.

In this early stage of agentic AI, many agents, while able to assist workers and perform certain tasks on their own, operate in isolation. An organization might have a marketing agent and a sales agent, but the two don't communicate with one another. Instead, the marketing agent aids marketers, while the sales agent aids salespeople, but their full potential is unrealized because there's little interoperability.

Interoperability, unlike reasoning, is not part of model training. Instead, it is a capability layered on top of models using data.

"The key part of agents is their infrastructure, and one main part of that is interoperability," Rajagopal said. "Agents working at scale, working across systems and platforms, is going to be super critical."

To facilitate interoperability, new protocols are being developed.

Model Context Protocol is an open standard framework introduced by Anthropic in late 2024 that standardizes how models share data with other systems and simplifies AI development for enterprises. As agentic AI gains momentum, MCP has gained support from tech giants AWS, Google Cloud and Microsoft, as well as data specialists including Boomi and ThoughtSpot.

More recently, the Agent2Agent protocol, which Google introduced in April, adds a standard for agents to communicate and collaborate, simplifying that aspect of AI development.

"Having standardized, secure ways for agents to interoperate is going to be super critical," Rajagopal said.

In addition, just as enterprises need data governance frameworks to simultaneously empower employees to work confidently with data while ensuring that data remains secure and its use does not violate regulations, agents need to be governed to ensure that they are operating within proper guidelines, he continued.

"You need to make sure that agents have access to the information that they are supposed to access so they are truly like a co-worker in how they interact with data," Rajagopal said.

Making it happen

Improving models and making agents more interoperable, while a clear path to improving agentic AI, is not simple.

One of the main hindrances is that models don't yet have enough memory and context, according to Leskovec. Humans can recall interactions, so they don't have to start over each time they discuss a given topic. Models, however, don't have that recall, at least not yet. As a result, their decision-making skills aren't as precise as required by many enterprises.

"That's one problem -- keeping context and memory from past interactions," Leskovec said.

Another hindrance is that models jump to conclusions without being interactive, he continued. Rather than asking questions to pinpoint the proper response to a given query, they respond to an initial query without seeking feedback.

For example, if an AI model were a waiter and a user said they were hungry, the model would bring them food without asking whether they have any dietary restrictions or what they want to order from the menu.

Meanwhile, because agentic systems are more complex than traditional machine learning models, it's more difficult to evaluate them. According to Sangani, the industry could use a standard for running agentic system evaluations. In addition, it's imperative that model developers emphasize safety and privacy when building agentic systems, he said.

"These agentic systems can be extremely powerful and can connect to numerous other systems and do tool calling, so building ingrained safety in the system is something we're taking very seriously," Sangani said. "Building in safety at the root is critical."

Ultimately, agentic AI's future hinges on collaboration between the research community and industry, some say.

For models to improve, they need the ingenuity of the research community combined with data from the enterprise world. For interoperability to improve, the same collaboration is needed with researchers advancing technology and industry providing the training material that enables agents to become more effective.

"Through my career at Stanford, we have worked with a lot of different organizations and found a lot of synergy where we can go forward in academic labs and do rigorous research to spin it out into the organization, where it really sees impact and makes a difference," Leskovec said.

In addition, because universities are independent of profit motive, they can assist with ethical questions about AI, he continued.

"We can step back and ask, 'How can this be truly solved to the benefit of humanity?' beyond trying to sell something," Leskovec said.

Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than 25 years of experience. He covers analytics and data management.