Bret Taylor (OpenAI chair, Sierra founder) says AI agents will be as essential as websites were in 1995. By 2025, you'll interact with brands mainly through their agents, not websites. He's bullish on Sierra (his company building customer-service agents for Sonos, SiriusXM), OpenAI (GPT-4o's voice is a breakthrough), and Harvey (legal AI agent). He warns that building a reliable, action-taking agent is much harder than a demo—AI is costly and unpredictable.
Bret Taylor discussed the past, present, and future of AI agents on the Invest Like the Best podcast. His core argument is that AI agents will fundamentally transform human-computer interaction and become a vital component of future technology. Drawing on his experience building Google Maps, serving
Bret Taylor (core developer of Google Maps, former CTO of Facebook, former co-CEO of Salesforce, Chairman of OpenAI's board, co-founder of Sierra) engages in an in-depth discussion with Patrick O'Shaughnessy on the current state and future of AI Agents. Taylor's core judgment is that AI Agents will become an essential digital asset for every brand, just like websites in 1995; by 2025, the primary way businesses interact with customers will shift from websites/apps to AI Agents, and interactions between Agents (personal Agents conversing with company Agents) will soon surpass human-to-Agent interactions.
Taylor argues that Agent is not a single concept but three distinct categories, each solving different problems.
Enterprise Agents (e.g., Sierra's product): Brand-owned, customer-facing conversational AI, analogous to websites in 1995. Taylor emphasizes: "In 2025, the way you interact with customers will be through your AI Agent." It must embody brand identity and handle the full spectrum of customer service, product recommendations, and more.
Persona-Based Agents: Internal-facing agents designed for specific job functions. Examples include software engineering agents (writing code), legal agents (e.g., Harvey), and operations analysis agents. Taylor notes that these agents are characterized as "narrow and deep"—focused use cases requiring deep domain expertise.
Personal Agents: Agents that work on behalf of individual users, such as planning vacations, managing calendars, and organizing inboxes. Taylor believes the biggest challenge for this category is "unlimited integration breadth"—different users rely on different calendar, email, and other systems.
Implication: Taylor and Sierra co-founder Clay have an internal bet—"When will traffic from personal agents to enterprise agents exceed traffic from real humans?" This suggests that inter-agent interaction will become a key future paradigm.
Taylor believes that the difficulty of building AI Agents is severely underestimated: building a demo is easy, but building an industrial-grade system is extremely difficult.
The core contradiction lies in the fundamental differences between traditional software and AI systems:
| Dimension | Traditional Software | AI Agent |
|---|---|---|
| Cost | Marginal cost approaches zero | Each inference incurs significant cost |
| Determinism | Same input → same output | Non-deterministic, difficult to reproduce |
| Speed | Millisecond-level | Relatively slow (Token generation takes time) |
| Core Advantage | Reliability, stability | Creativity, flexibility |
Taylor proposes a "Goals and Guardrails" framework to replace the traditional "rule-based system" mindset. He shares Sierra's approach: using a multi-layer architecture where "one AI model supervises another AI model" to improve statistical reliability.
Key Data: Taylor points out that "90% system accuracy" has vastly different implications across scenarios — it may be acceptable for consumer applications, but "if it involves revenue or compliance, a 10% error rate is absolutely unacceptable."
Falsification Condition: If the non-deterministic nature of AI models cannot be effectively controlled through engineering methods (such as multi-model supervision), the deployment of Agents in serious business scenarios will be limited.
Taylor proposes a core formula for agent capability: Factual Knowledge + Procedural Knowledge + System Integration = An Actionable Agent.
Factual Knowledge: Through RAG (Retrieval-Augmented Generation) technology, the agent responds based on the company’s knowledge base rather than relying on model training data. Taylor believes this is "necessary, but far from sufficient."
Procedural Knowledge: This is the element Taylor considers the most underestimated. "If a Sonos speaker breaks, what would the best Sonos engineer ask you and do to diagnose it?" This includes the best sales processes, the best customer service workflows, the most effective product description methods, and more.
System Integration: The agent must be able to take real action—processing returns, changing subscriptions, sending signals to satellites to refresh radios, etc.
Data Support: Sierra’s customer onboarding cycle is 1–3 months, employing a high-touch model ("IKEA, not Home Depot") to ensure that even non-AI experts can deploy it.
Case Study: For one of Sierra’s early clients (a footwear company), the agent was originally intended to handle customer service queries like "Where is my order?" However, the very first real conversation was: "I’m going to a wedding in Hawaii—what sandals go with my bridesmaid dress?" Taylor uses this to illustrate that the "free-form" nature of agents can lead to use cases beyond expectations.
Taylor holds a dual view on AI’s effect on social inequality: optimistic about lowering barriers to entry, but wary of accelerating the digital divide.
Optimistic side: AI lowers the "gatekeeping" in professional fields. "Think about how many obstacles James Cameron had to overcome to make The Terminator or Christopher Nolan to make Inception?" Taylor believes that people with taste and judgment can now create high-quality work with fewer resources—"requiring less social permission and financing."
Wary side: Taylor acknowledges that "the best software engineers will become far more leveraged," shifting from "someone who types code on a keyboard" to "an operator of a code-generating machine." He specifically highlights a key difference: previous technological revolutions (PCs, smartphones, broadband) required years of infrastructure investment, whereas AI is a "software-defined revolution," and the pace of change could be much faster. "The transition from an industrial economy to a service economy took decades, but changing the nature of work may take only a few years."
Taylor’s clear judgment: "Society will reshape the economy around the technologies we create. We are status-seeking creatures and will build new economies around the technologies that define society—but this process could be very uncomfortable."
Taylor believes that GPT-4o's voice conversation capability marks "the first time a sci-fi movie-style human-computer interaction has truly felt right."
He traces the evolution of human-computer interaction: punch cards → mouse and keyboard → touchscreens → conversation. "My grandparents skipped the PC era but used the iPad. Now imagine if they were still here—they would just need to speak. Everyone knows how to speak."
Core thesis: Multimodal models (text + image + video + audio) will allow "technology to fade into the background." Taylor says: "We spend too much time staring at screens, and that's bad for our relationships. I want technology to become more powerful while simultaneously disappearing from what we do."
Extrapolation: Taylor believes that smartphones "will remain the dominant interface for the foreseeable future," not because they are the best, but because they "already exist and are extremely mature." However, he remains open to new interaction devices (smart speakers, AirPods, brain-computer interfaces)—"Silicon Valley was born for this moment."
| Position | Guest Stance | Key Data |
|---|---|---|
| Sierra | Bullish (on the company) | Customer onboarding cycle 1–3 months; uses 8–9 models; customers include Sonos, SiriusXM, Weight Watchers |
| OpenAI | Bullish (serves as Board Chair) | Nonprofit structure, mission "to ensure that AGI benefits all of humanity"; GPT-4o described by Taylor as a "breakthrough interaction experience" |
| Harvey | Bullish (mentioned) | Legal AI Agent, no specific data provided |
| Sonos | Neutral (mentioned as a Sierra customer) | Uses the Sierra platform to build a customer support Agent |
| SiriusXM | Neutral (mentioned as a Sierra customer) | Agent named "Harmony" can send signals to satellites to refresh radios |
| Weight Watchers | Neutral (mentioned as a Sierra customer) | Offers a 24/7 AI coach within the App |
1. Taylor’s definition of an Agent: "The word Agent comes from agency (autonomy), essentially a system capable of autonomous reasoning and action." — Emphasizing autonomy rather than simple conversational ability.
2. "90% accuracy is an inkblot test": Taylor points out that 90% accuracy means entirely different things in different contexts—acceptable for consumer applications, but "when it involves revenue or compliance, a 10% error rate is absolutely unacceptable." This is the key dividing line for determining whether an Agent is ready.
3. "From rule-based systems to goal + guardrail systems": Taylor believes this is the biggest paradigm shift in software engineering in the AI era. Traditional software pursues "same input → same output," while AI systems must embrace non-determinism, replacing the framework with "goal + guardrails."
4. "Conversations between personal Agents and corporate Agents will soon outnumber those between humans and Agents": Taylor has an internal bet with co-founder Clay, reflecting his strong conviction that inter-Agent interaction will become the mainstream paradigm.
5. "The speed of the AI revolution is defined by software": Taylor notes that past technological revolutions required years of infrastructure investment, whereas AI adoption can "happen as instantly as the technology becomes effective" — this is his primary concern regarding society’s ability to adapt.
6. "Every company should ask: What job do customers hire us to do?": Taylor cites Clayton Christensen’s "Jobs to be Done" framework, arguing that in the AI era, the biggest risk for companies is "confusing 'how' with 'what value is delivered.'"
7. "Sheryl Sandberg taught me: Don’t try to make a new job fit you, but adapt yourself to the new job": Taylor shares a pivotal turning point when he became Facebook’s CTO at age 29 — "maintaining a loose sense of identity" allowed him to transition from a technical expert to a leader.
8. "Sierra’s values are intensity and craftsmanship": Taylor believes that in the AI frenzy, concepts are obvious, and "the key is execution — Amazon wrote the history books, not buy.com."