From Answering Questions to Taking Action: The Shift of Competitive Moats in the Age of AI Agents

Event

In early May 2026, Google beta-tested an AI agent called "Remy," while Meta simultaneously developed a similar tool codenamed "Hatch." These moves came after OpenClaw had gained widespread market acceptance, with NVIDIA CEO Jensen Huang even publicly stating that "every company needs an OpenClaw strategy." According to leaked Google internal documents, Remy is positioned as an "always-on personal agent" capable of deep integration with Gmail, Chrome, Calendar, and other services, proactively monitoring user behavior and executing tasks on their behalf.

Observation

This competition is not patching on top of ChatGPT—it's a qualitative shift. Over the past five years, the value proposition of chatbots has been "finding answers faster." Users still needed to judge, verify, and decide. But the agent era flips this model: the system doesn't just answer "what should I do," it directly "does it for you."

This means three fundamental changes across different dimensions.

First, the depth of data control. Information providers see queries; agents see decision chains. When Remy monitors your emails, schedule, and searches daily, it accumulates not "what questions this person asked," but "what choices this person made in what contexts, and how they turned out." This decision-history data is more authentic and predictive than any survey.

Second, switching costs and stickiness. You can easily switch from ChatGPT to Claude because both are just text interfaces. But if Remy has already handled three months of your email filing, schedule coordination, and expense reimbursement processes, switching costs skyrocket—because the new agent needs to relearn your preferences and habits. Stickiness leaps from "tool" to "habit."

Third, platform control. AI that answers questions is a tool; AI that takes action is a decision-maker. It determines which emails you should see, which meetings you should attend, which products you should buy. This means Google and Meta are no longer search engines or social platforms, but "agents" that directly influence user behavior—and agents possess absolute advantage in information asymmetry.

Pattern

This is a pattern that has repeated every 30 years: value chains shift downward from the information layer to the decision layer, and monopoly power strengthens accordingly.

Microsoft upgraded from a software company to a "workflow orchestrator"—Office 365 is no longer just word processing, but decides for enterprises who should receive what information and when to archive it. Google upgraded from a search engine to an "ad decision engine," not just answering "which restaurant should I go to," but directly deciding which restaurant's ad you see. Now AI agents are doing the same thing: upgrading from "helping you think" to "helping you do."

Principle

When capability upgrades shift the center of gravity from information provision to action execution, they trigger structural reorganization of the market. Platforms that control "what to do" have stronger monopoly power, greater value, and deeper moats than platforms that control "what to know."

Application

For pure model companies like OpenAI and Anthropic, this is a warning signal. If the moat of AI agents lies in "integration depth" rather than "model intelligence," their best position is to be integrated into Google/Meta's ecosystem rather than compete independently. For Chinese tech companies, being locked into the information layer (search, Q&A) is no longer enough; they must rapidly move into the action layer (e-commerce decisions, payments, integrated calendar management agents) to maintain influence in the next round of competition.

Counterargument

Some will say: the smarter the agent, the more easily users can be manipulated, and trust crises will destroy them. But history tells us this concern is often overstated. People once worried Google would use search results to manipulate elections; Facebook also faced trust crises. What happened? The compound effect of switching costs and network effects ultimately outweighed ethical concerns. Agents will be the same: as long as they get it right 90% of the time, users will trust them with even more.