Last year, Salesforce and HubSpot announced they were all in on AI agents. Salesforce CEO Marc Benioff said the company was doing a “hard pivot” and would be 100% about its Agentforce platform.
“[Agentforce] is the third wave of AI — advancing beyond copilots to a new era of highly accurate, low-hallucination intelligent agents that actively drive customer success,” he said at Dreamforce last September.
So, what are agents?
Here’s what the marketing materials say: Agents are AI that use tools to accomplish goals. They are intelligent systems that perform tasks without human intervention. Agents are characterized by their ability to operate independently, make decisions, and take action.
“Agents are nothing new,” said Christopher Penn, TrustInsight’s co-founder and chief data scientist.
“[Agents] aren’t really agents,” said Paul Roetzer, CEO of The AI Marketing Institute.
Dig deeper: Watch an excerpt of MarTech’s conversation with Christopher Penn
So, what are they?
Penn sees them as self-driving apps. “They have a language model (LLM) embedded in part of them, plus some infrastructure around them that allows them to run on a certain schedule or have more than one language model working in tandem,” he said.
Dig deeper: What’s the difference between agentic AI and generative AI?
Roetzer thinks that is giving them too much credit.
“They just changed the name of bots,” said Roetzer. Agents were supposed to be much more impressive than what we have now.
“We were promised something that seamlessly uses your computer, completing tasks like filling forms and booking travel arrangements with access to your accounts,” he said.
To Roetzer, the current reality is that although they utilize LLMs in their processes, they lack the comprehensive capabilities that would differentiate them from bots.
“Because there’s an LLM used in the process, because AI is being used in the steps, it feels like everyone just became OK with calling it an AI agent,” he said.
Penn essentially agrees with this. He said agents are apps with an embedded LLM and infrastructure that allows them to run on a specific schedule and then hand the product off to another app that does something with it. The addition of the LLM is an interesting addition to an existing technology. However, “this is not new. Zapier as a company has basically had agents” for around a decade.
Dig deeper: Salesforce Agentforce: What you need to know
“Microsoft’s AutoGen is an agent framework that allows you to have one model write something,” he said, “And then the second model, in this controlled sequence, proofreads it and says that you didn’t follow the assignment.”
These caveats don’t mean these “agents” aren’t incredibly useful. As Penn points out, these things are powerful and straightforward in what you can do with them. As a result, they have seemingly
an infinite number of use cases.”
“I suspect a very expensive consultant somewhere wanted to increase his billing rate and, came up with agentic AI, as opposed to, just apps,” he said.
You will know a real agent because it won’t need human interaction.
“But, for the foreseeable future, the human’s likely going to be heavily involved in most agent applications,” Roetzer said. “It seems like this year is going to be a lot of people doing a lot of custom-builds of agents that are basically a bunch of if/then statements mashed together with some LLMs.”
To be clear, real autonomous agents are coming. Improved reasoning capabilities will eventually let AI go through a chain of thought. That, and improved multimodal models that can produce text, images, video, and audio, means agents will ultimately be more than we were originally promised.