May 2, 2025 · 36m · a16z
What Is an AI Agent?
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the a16z podcast, partners Guido Appenzeller, Yoko Li, and Matt Bornstein analyze the technical definitions, architectural frameworks, and economic realities of AI agents. They critique market hype around human job replacement, providing strategic guidance on enterprise software pricing, defensible moats, and the long-term normalization of AI.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 0.0 Guest teaching 2.3 Guest disagreement 2.5 The host pushing back 0.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 6:21 Matt rejects Anthropic's agent definition
Matt Bornstein directly challenges Anthropic's framing of agents as LLMs in loops with tools, pointing out that basic ChatGPT web search interactions already meet that exact definition.
Hardest push from the host ▶ 14:16 Guido pushes back on treating agents as distinct from functionsActing as discussion leader, Guido Appenzeller challenges Yoko Li's definition of low-level agents, pressing whether an external caller can distinguish an agent from a standard software function.
Biggest teaching moment ▶ 23:44 Yoko illustrates application layer pricing moatsYoko Li educates the panel on value capture by demonstrating how Pokemon Go charges thousands of times raw cloud storage costs for a simple JSON blob due to its application monopoly.
The host holds their own ▶ 25:58 Guido breaks down the architectural systems view of agentsGuido Appenzeller demonstrates deep technical authority by detailing how agent architectures decouple heavy GPU model calls and external state management from lightweight local loops.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Defining AI Agents Across Technical and Marketing Perspectives | 0 | 2 | 4 | 0 | The host does not participate in the dialogue. Guest Matt Bornstein adopts a contrarian stance, dismissing the term agent as largely a marketing rebrand for standard AI applications rather than a true decade-long technical breakthrough. | |
| Exploring Degrees of Agentic Behavior and Copilots | 0 | 2 | 4 | 0 | Host involvement remains non-existent. Matt Bornstein challenges Anthropic's definition of an agent as an LLM in a loop with tool use, arguing that standard ChatGPT reasoning queries already meet those criteria. | |
| Evaluating Human Replacement and Work Automation | 0 | 3 | 4 | 0 | The host is passive while guests debate whether AI replaces human workers. Matt Bornstein argues forcefully against human replacement narratives, claiming human work fundamentally requires intent and creative decision-making. | |
| Pricing Models for AI Agents | 0 | 2 | 2 | 0 | The conversation is collaborative as the guests explore software pricing models. Yoko Li explains the distinction between per-seat pricing for human users and usage-based pricing for machine services. | |
| Identifying True Enterprise Value and Moats | 0 | 3 | 1 | 0 | The guests maintain a cooperative tone while discussing enterprise moats. Yoko Li uses Pokemon Go bag storage as an example of application-layer monopolies commanding massive margins above raw cloud infrastructure costs. | |
| Architectural Systems View of Agent Capabilities | 0 | 2 | 2 | 0 | Guido Appenzeller outlines the modular architecture of AI agents. Matt Bornstein points out that handling non-deterministic LLM output inside program control flows remains a major unsolved engineering challenge. | |
| Data Silos, Walled Gardens, and Web Scraping Challenges | 0 | 2 | 2 | 0 | The discussion covers data silos and scraping barriers. Guido Appenzeller shares an anecdote of a deep research AI model attempting to bypass captcha mechanisms as part of its internal reasoning step. | |
| Future Vision and the Normalization of AI | 0 | 2 | 1 | 0 | The segment offers predictions for AI integration. Matt Bornstein cites research suggesting AI should be viewed as a normal general-purpose technology like electricity rather than a utopian or dystopian force. |
Statements from this episode (23)
Li: Reasoning and decision-making define all AI agent use cases
“I almost feel like for all the use cases with Describe, there's one element that all agents have, which is reasoning and decision.”
Bornstein: 'AI agent' is just a rebrand for standard AI applications
“I kind of think agent is just a word for AI applications, right? Anything that uses AI kind of can be an agent.”
Bornstein: Most current commercial AI agents are just weekend demos
“Most of what we're seeing in the market now is like, is not the decade version of this problem. It's like the weekend demo version of this problem.”
Bornstein: AI systems cannot be defined by user inputs because LLMs accept anything
“I just think it's really tough To define a system based on what someone says to it, right? Because these are by design, unstructured inputs, like these systems will accept literally anything. And so sure, if you tell it, you know, what's today's weather, I wou…”
Appenzeller: AI user interfaces are specializing into interactive copilots and async plugins
“It seems like we're seeing to some degree a specialization of user interfaces in sort of two directions, right? There's, let's say, a cursor or something like that, which really emphasizes the tight loop between the user, the tight feedback loop between the us…”
Li: AI agents are multi-step LLM chains with decision trees
“I actually feel like it's a multi-step LM chain with a decision tree.”
Li: AI slows net new human hiring rather than firing existing workers
“But we are seeing headcount growth in some areas are slowing. So it's not that existing jobs are being replaced. It's more like they're hiring net new. Humans slower.”
Appenzeller: AI will enable one worker to replace two, rather than direct human elimination
“I mean, I think in few cases, humans will get replaced by AI. In most cases, you know, two humans will get replaced, one human that is more, by one human that's more productive with AI.”
Li: Amazon Go relied on human data labelers despite automated computer vision claims
“There's an Amazon Go supermarket a while back in Soma. So like they, I think they were advertising that it's computer vision models behind the scenes. Identifying what you took from the supermarket. But then people found that they hire a lot of people behind t…”
Bornstein: Enterprise buyers force AI vendors into GPU cost-plus pricing
“In practice, I think most buyers are actually pretty sophisticated about what's going on under the hood. And to your point, they know it's like pretty simple stuff happening. And so what does it cost you to run all these GPUs? And we'll pay you some premium ov…”
Yoko Li: Infrastructure pricing splits per-seat for humans, usage-based for machines
“And traditionally for infra, a rule of thumb, like not always the case is that if the surface is used by a human, it's a per seat pricing. And if it's a service that's used by other machines, it's a usage based pricing.”
Bornstein: Most AI startups do not understand their value creation yet
“And the reality is most AI companies don't know what value they're generating yet. Like this is like so new and so nascent that it's like, hey, we're just going to charge something that we're not going to lose money on.”
Bornstein: AI coding tool prices are decoupling from underlying model costs
“Code, you can somewhat see the decoupling of price from the underlying technology now, because there's very clear, like it really works. There's very clear ROI to all of the to people who use it.”
Yoko Li: Pokémon GO charges thousands of times underlying storage costs
“And as an infrastructure investor, I invest in storage businesses. And then when I look at how much I need to pay for like, 30 extra Pokemon, it was thousands of times more expensive than what storage is.”
Yoko Li: AI agent companions will monetize virtual assets like games
“Yeah, and then we'll see the agent, AI agent version of this. I can't wait to see the AI companion version of this. Paying storage for AI companions wardrobe.”
Appenzeller: Building AI agents architecturally mirrors traditional SaaS software
“And I personally think that architecturally There really is no difference between your typical SaaS software today and Agent in terms of how you build it, right?”
Li: Specialist AI startups will win over foundation model creators
“I actually think the winners will be the specialists, not the foundational models. It's the people who will build on top of the foundational models or fine tune the foundational models.”
Bornstein: AI data access will be solved by scraping, not protocols
“These problems are rarely solved by defining a new protocol and just saying like, Hey, if we make it easy for people to give away their core assets, they'll just do it. You know, obviously, you know, that's very unlikely to work, but someone eventually will so…”
Yoko Li: Web browsing with AI agents remains slow and unreliable
“Today, web browsing using an agent doesn't work super well. It's very slow. It's very clunky. You have to try it multiple times for it to do any task.”
Appenzeller: Consumer websites are deploying anti-agent CAPTCHAs against AI
“All the consumer sites are starting with more and more complex anti-agent captures trying to keep out their agents because they only want the humans that have attention to come to those sites.”
Appenzeller: AI agents will use most human accessible tools in two years
“Look, I think that the positive vision is that in two years, we figured out how an agent working on my behalf can use most of the tools that I have access to”
Li: Multimodal training will unlock net new AI agent capabilities
“So I will actually bet on multimodality when it comes to, like, if we train more model train, like, the model with, like, Different traces of like clicking on buttons on the website, navigating the web using different devices, drawing producing vector art. I t…”