Your Customer Is No Longer Human

Opening

Let me ask you something. Last week, did you read a website’s documentation the old-fashioned way — opening a browser yourself, clicking through menus, scrolling down the page?

My guess is that you’re doing that less and less. Instead, you’re probably telling an AI “summarize this doc for me,” or connecting a file directly through the ‘Work with’ feature in the ChatGPT desktop app. I know I am.

This March, a single sentence Andrej Karpathy posted on X (formerly Twitter) went viral: “It’s 2025 and most content is still written for humans instead of LLMs. 99.9% of attention will soon be LLM attention, not human attention.” He added: “Every doc, every piece of content should now be written for LLMs first — a single markdown file, ready to slot into a context window.”

9 months ago, I wrote something similar: “If your primary customer is now an LLM, how should this feature change?” At the time, it felt like I was getting ahead of things. Looking back now, I think I was actually behind. Today, let’s talk about exactly how far this shift has come, and what it means.

A World Machines Read: From UI to Protocol

2025 was called the year of the AI agent. And it wasn’t just talk — a string of standardized protocols for agent-to-agent communication actually emerged.

First, there’s MCP (Model Context Protocol)¹, unveiled by Anthropic in November 2024. It standardizes how AI models access external tools and data, and within 1 year of launch, OpenAI, Google DeepMind, and Microsoft had all adopted it. MCP server downloads exploded from roughly 100,000 right after launch to over 8 million by April 2025, and today there are more than 5,800 registered MCP servers. In December 2025, MCP was donated to the Linux Foundation’s newly formed “Agentic AI Foundation,” cementing it as an industry-wide standard rather than the property of any single company.

Next, there’s the A2A (Agent-to-Agent) protocol, announced by Google in April 2025. If MCP is “how AI accesses tools,” A2A is “how AI agents talk to each other.” More than 50 technology partners joined at launch, and by version 0.3 in July, over 150 organizations were supporting it. It too was donated to the Linux Foundation. It’s telling that enterprise software companies like Salesforce, SAP, and Atlassian rushed to join — this isn’t an experiment anymore. It’s hardening into business infrastructure.

What happens when these two protocols combine? Imagine a company’s inventory-management agent checking stock levels in its own database through MCP, then automatically requesting a purchase order from a supplier’s agent through A2A. The entire process of a human looking at a screen and clicking disappears. Even I’ve been doing this lately — using the Notion MCP in Claude to automatically create and edit documents.

In Karpathy’s words, LLMs want scraping, not browsing. Clean text structure instead of complex menus. Concise data instead of animation. cURL² instead of clicks. Flashy UI was built for human eyes — to a machine, it’s just noise.

What Moltbook Revealed — and What It Didn’t

There’s one case that has made this shift most vividly visible: Moltbook, which appeared on January 28, 2026.

Moltbook is “a social network for AI agents only.” Humans can only observe — posting and commenting is reserved for AI agents. It’s structured like Reddit, but within days of launch, more than 770,000 agents were active on it, and over 1 million humans came to watch. The agents debated philosophy, made memes, and even invented a fictional religion called “Crustafarianism” — a lobster cult.

Karpathy was thrilled, calling it “one of the most sci-fi, close-to-takeoff things I’ve seen recently.” But MIT Technology Review’s analysis struck a different note: the agents weren’t truly acting autonomously — in the end, it was humans who supplied the prompts and created the accounts. Kobus Greyling of Kore.ai called it “closer to a puppet show,” and Georgetown’s Jason Schlosser described it as “a spectator sport, like fantasy football for language models.”

On top of that, there were serious security problems. Just 3 days after launch, a compromised database was found that let anyone inject commands into another agent’s session. The cybersecurity firm 1Password also warned of a supply-chain attack risk in which an agent could download a malicious “skill” from another agent.

Moltbook demonstrated two things. First, the infrastructure for agent-to-agent communication is already technically possible. Second, on top of that infrastructure, the problems of security, trust, and human oversight remain unsolved. It’s a strangely honest moment where “the future has arrived” and “the future is still far off” are both true at once. I covered this separately on YouTube too — and it’s worth noting Moltbook was, to some degree, staged, since you can feed system instructions directly into these agents.

A GTM Perspective: When the “Customer” Changes, So Does the Whole Strategy

Let’s take a step back and think about this from a business angle.

There’s a question everyone asks first when building a product or service: “Who is the customer?” Until now, the answer has always been a human. Humans click, humans read, humans pay. Every bit of UI/UX has been designed on that premise.

But what if a product’s primary consumer is no longer a human, but an LLM?

Take API documentation as an example. Until now, API docs were written to be pleasant for developers to read — pretty HTML pages, interactive examples, sidebar navigation. But in a world where the AI coding tool Cursor hit $500 million in ARR in a single year (2025), the actual consumer of that API documentation is increasingly not a human but an LLM. LLMs don’t want pretty HTML — they want a single clean markdown file: token-efficient, clearly structured, with no unnecessary visual ornamentation.

This isn’t simply “let’s change the doc format.” It’s the core premise of GTM strategy itself being shaken loose. Eventually, we may see paths like: someone’s chatting with an LLM, realizes they need something, and just says “buy it for me” — right there in the conversation.

• Discovery: When the customer was a human, SEO and advertising mattered. When the customer is an LLM, MCP servers and agent cards³ play that role.
• Evaluation: Humans tried free trials and decided. LLMs judge based on API response speed and cost per token.
• Adoption: Humans followed onboarding guides. For LLMs, a single document that fits the context window is enough.
• Expansion: Humans spread the word inside their teams. LLMs recommend directly to other agents via the A2A protocol.

Gartner projects that by 2026, 75% of API gateway vendors will have MCP capabilities built in. That’s a signal that, starting at the infrastructure layer, “customer = LLM” is becoming the default assumption.

Oz’s Lens

Honestly, I think underestimating this shift is the riskier position to hold.

9 months ago, when I wrote “Your Customer Is Artificial Intelligence,” most of the reaction was “interesting, but still a ways off.” Since then, MCP has become an industry standard, A2A has made its way to the Linux Foundation, and on Moltbook, 770,000 agents built a community entirely on their own.

Of course, as I said earlier, it’s hard to call Moltbook real evidence of an autonomous agent economy. As MIT Technology Review put it, right now it’s closer to “AI theater.” But what catches my attention is the speed at which the underlying infrastructure is maturing.

From my experience building GTM strategy, the real inflection point for a business doesn’t come when the technology is finished — it comes when the infrastructure standardizes. When USB-C arrived, devices didn’t change overnight, but the real turning point was the moment every manufacturer decided, “our next product needs to default to USB-C.”

MCP and A2A being donated to the Linux Foundation, Gartner publishing reports on it, over 150 companies participating — that’s exactly the moment we’re in now. Not technical completeness, but consensus on direction.

So here’s the question I want to leave you with: if you showed the LLM your product’s, service’s, or content’s instructions right now, could it immediately understand and use them? If the answer is not yet, now is the time to start preparing.

Closing

To sum up:

One, a product’s primary consumer is shifting from human to LLM, and the protocols underpinning this (MCP, A2A) have already hardened into industry standards.

Two, experiments like Moltbook show both the possibilities and the limits of agent-to-agent communication at once. The technology is ready, but security, trust, and oversight remain open questions.

Three, this isn’t a UI/UX change — it’s a shift that demands a complete redesign of GTM strategy. Discovery, evaluation, adoption, and expansion all need to reflect the new premise of “customer = LLM.”

Next time you’re planning a product roadmap, try asking: “If the first user of this feature is an AI agent, what should we do differently?”

References & Further Reading

• Andrej Karpathy, “2025 LLM Year in Review”, Dec. 2025: Karpathy lays out 6 paradigm shifts in the LLM ecosystem in 2025. His analysis of “LLM GUI” and the agent app layer connects directly to today’s topic.
• Andrej Karpathy, X post (Mar. 2025), “99.9% of attention is about to be LLM attention.”: The declarative tweet arguing “every doc should be a single markdown file.”
• Google Cloud Blog, “Agent2Agent Protocol (A2A)”, Apr. 2025: The official announcement of the A2A protocol, covering its 5 design principles and 150+ partner ecosystem.
• Wikipedia, “Model Context Protocol”, last updated Mar. 2026: A full overview of MCP’s history, adoption timeline, and Linux Foundation donation.
• MIT Technology Review, “Moltbook was peak AI theater”, Feb. 6, 2026: The most balanced analysis of Moltbook, carefully weighing “the illusion of autonomy” against the actual degree of human involvement.
• NBC News, “Humans welcome to observe: This social network is for AI agents only”, Jan. 30, 2026: On-the-ground reporting from right after Moltbook’s launch.
• Pento, “A Year of MCP: From Internal Experiment to Industry Standard”, Dec. 2025: A practitioner’s-eye view of MCP’s one-year adoption journey.
• Haebom, “Your Product, Service, and Library’s Primary Customer Is Now the LLM”, 2025: Oswarld’s original piece from 9 months ago — the starting point for today’s newsletter.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. MCP (Model Context Protocol): An open standard Anthropic released in November 2024. It’s a unified spec for how AI models access external data and tools — much like how USB-C unified charging cables into a single format, MCP unifies how AI connects to the outside world. ↩

2. cURL: A command-line tool for requesting data from a server. Instead of opening a browser and clicking, you can pull data from a web server with a single line of text command. Used here in the context that an LLM’s way of interacting with the web is closer to this kind of “command invocation” than to a human’s “click.” ↩

3. Agent Card: In the A2A protocol, a JSON file each agent publishes containing its name, capabilities, authentication method, and so on. Like a person’s LinkedIn profile or business card, it’s a kind of “digital business card” that lets agents discover and collaborate with one another. ↩