Buzzwords that are stretched to the point of being meaningless are as old as the tech industry itself. The top current one is “AI agent” and its variants, like “agentic.”
So, unsurprisingly, no one really knows what an AI agent is. Even people with software engineering backgrounds who work for Andreessen Horowitz, one of the premier venture capital firms madly funding AI startups, say that there’s no agreed-upon definition.
Three a16z infrastructure investment partners — Guido Appenzeller, Matt Bornstein, and Yoko Li — tried to come up with their own definition of agent during a recent podcast episode called “What Is an AI Agent?”
For perspective, a16z, backer of such hot AI companies as OpenAI and Anysphere (maker of Cursor), is so gung-ho on the AI opportunity that it’s reportedly attempting to raise a $20 billion mega-fund to invest even more heavily in the sector, sources told Reuters last month. Back in September, two other a16z VCs explained the firm’s excitement, writing on its corporate blog: “We believe every white-collar role will have an AI copilot. Some of these roles will be fully automated with AI agents.”
To cash in on the buzz, “a continuum” of AI startups are describing their products as agents, Appenzeller says.
“The simplest thing that I’ve heard being called an agent is basically just a clever prompt on top of some kind of knowledge base,” he said. This so-called agent takes a question from a human, then fetches a “canned” response, such as with IT help desk support.
But lately, companies that make agents, or want to make them, have been describing them as human worker replacements.
To really do that, their AI software would have to be “something close to AGI,” Appenzeller says, which means “it needs to persist over long periods of time” and “it needs to work independently on problems.”
Yet such a thing “doesn’t work yet,” both he and Li said.
The reality is that getting this nascent AI agent tech to work reliably well has been a surprisingly hard journey, Jaspar Carmichael-Jack, the CEO of sales AI agent company Artisan, told TechCrunch last month. Carmichael-Jack is still hiring humans, despite his startup’s viral “stop hiring humans” ad campaign.
For an AI to become a true human worker replacement, there are significant technical issues to solve, such as persistent long-term memory (and costs associated with that), and stomping out hallucinations. Because no company wants to hire an employee — human or artificial — who can’t remember a previous conversation and who also randomly lies.
During the podcast, the a16z trio did land on a solid definition of what’s possible today. As Li described, an AI agent is a reasoning, multi-step LLM with a dynamic decision tree.
In other words, she said, an agent isn’t a bot that just does a task when asked; it must also be able to make decisions about the task and take action autonomously, like grab a list of prospects from a database, decide which ones to email, and write the emails. Or write code and decide where to insert it.
As for whether agents could actually replace humans in the foreseeable future, all three VCs agreed they could be used to handle some tasks humans do now, just like automation has always done. But this may actually lead to companies hiring more human workers, not fewer, as productivity rises.
Bornstein said he can’t envision a time — given the current state of agents — when humans will be unnecessary. From “our perch in Silicon Valley,” the tech industry can “forget” that most people have jobs that require human creativity and “thinking,” he described. To replace humans with a bot, “I’m just not sure that even is kind-of theoretically possible,” he said.
Still, such human replacement rhetoric — often done for marketing/business model and/or pricing reasons — is “a big reason for the confusion we’re experiencing now,” Bornstein says.
The upshot is, if those seeing all the most cutting-edge uses of AI agents are skeptical about the boldest claims AI agent companies are making today, that’s probably a good sign the rest of us should be, too.
