AI Agent Confusion Starts With the Word Itself

In Silicon Valley, no label has caught on faster or become more nebulous than “AI agent.” The term is now affixed to products ranging from LLM chat interfaces to autonomous workflow systems, and its ambiguity is quickly becoming part of its appeal. Yet behind the marketing sheen, some of the industry’s leading voices are starting to ask a harder question: what are we really talking about?

“If you look closely, most of what’s being branded as agents today is pretty thin,” said Guido Appenzeller, a general partner at Andreessen Horowitz. “It might just be a clever prompt stacked on top of a knowledge base.” In other words: what passes for an “agent” in 2025 might, in a different year, just be called software.

Appenzeller’s take is blunt, but not isolated. On a recent episode of the firm’s podcast, his colleagues Matt Bornstein and Yoko Li joined him in unpacking the growing disconnect between the term’s use and its substance. From vague product demos to inflated pricing models, their conclusion was clear: most agents aren’t actually agentic.

The Elastic Definition Problem

A recurring theme in the discussion was the lack of definitional rigor. “I don’t think anything we have are actually agents,” said Bornstein, who pointed to the overuse of the term across sales decks and investor updates. “It’s a word that means too many things to too many people.”

The problem, they argued, stems from the continuum of what constitutes agentic behavior. At one extreme, an agent could be little more than a prompt-template interacting with an LLM. At the other, it’s a persistent, self-improving digital entity capable of long-term planning and tool orchestration.

Some vendors have leaned hard into the latter. Not because they’ve built it, but because it sells well. “Startups are saying: we’re replacing a $50,000 human with a $30,000 agent,” Appenzeller said. “That narrative works in early-stage sales. But it doesn’t hold up if you’re just routing responses from a chatbot.”

From Tool to Worker? Not Quite.

To some degree, the agent label has become a pricing strategy. If co-pilots are assistive tools, agents are marketed as autonomous replacements. But in reality, that leap is rarely justified. Most so-called agents today are still dependent on user inputs, predefined logic, or brittle chains of API calls.

“We’re not seeing full replacement of humans,” Li noted. “It’s more like companies are hiring slower because workers are becoming more productive with AI.” Bornstein added: “In many cases, it’s just one person doing the work of two—with better tools.”

This raises a deeper question: if agents aren’t replacing humans, are they just new interfaces for existing systems?

In some respects, yes. Much of the functionality attributed to agents searching a knowledge base, triggering workflows, interacting with external APIs can already be achieved with traditional software patterns. What’s new is the LLM at the center: adding flexible language understanding, probabilistic decision-making, and a sense of “conversation.”

That’s also where the complications begin.

A Loop of Uncertainty

One proposed technical definition of agents, an LLM operating in a loop with tool use sounds simple. But the real-world implementation is anything but.

“Integrating the output of an LLM into your control flow isn’t trivial,” Bornstein explained. “It introduces non-determinism. You’re not just calling a function—you’re incorporating the judgment of a statistical model into the logic of your system.”

As more software attempts to include “autonomous steps” powered by LLMs, managing uncertainty becomes a key architectural challenge. Traditional SaaS was built on determinism; agents, as envisioned, are anything but.

Co-Pilots vs. Executors

Interestingly, the agent conversation is also splitting along UX lines. On one end are systems that work in tandem with users, copilots, chat-based tools, interactive notebooks. On the other are back-end agents that are handed a task and asked to complete it without further human input.

Both models serve distinct purposes. But conflating them under a single banner only adds to the confusion.

As Bornstein put it: “It’s really hard to define a system based on what someone says to it.” A prompt like “translate this to JSON” is a tool. A prompt like “triage these emails and draft responses” feels more like an agent—but the underlying system might be almost identical.

When the Sales Pitch Drives the Stack

Ultimately, the trio argued, the “agent” label is being driven as much by sales and pricing incentives as by real technical distinctions.

“There’s this sense that if you call it an agent, you can price it like labor,” Appenzeller said. “But over time, pricing in software always converges to marginal cost. And for most AI agents today, that cost is pretty low.”

That shift is already underway. As buyers get more sophisticated, they’re asking tougher questions: What’s really under the hood? Is this reusable logic or a brittle demo? Are the productivity gains real, or is this just window dressing?

The firm believes that the long-term winners in AI won’t be the ones who coin the buzziest labels but those who solve real problems with clear interfaces and reliable infrastructure. “If we stop using the word ‘agent’ entirely in five years, I’d consider that a win,” Bornstein said.

What Comes Next

In time, the concept of an agent may mature into a standardized software pattern. Or it may dissolve into the fabric of everyday apps, like APIs and cloud infrastructure before it.

But for now, the burden of clarity falls on builders and vendors. “Let’s talk about what the system actually does,” said Appenzeller. “Where is it deployed? Why does it matter? That’s a more honest conversation than just calling everything an agent.”