I will die on this hill: the general-purpose AI agent is a demo, not a product. The horizontal "agent that does everything" is going to be the next-generation chatbot — fun at the launch keynote, abandoned in production, the cause of an expensive postmortem eighteen months later. The future of agentic AI is vertical. The proof is in the revenue and the renewal rate, and in the founders I have personally seen hit the same wall.
I have watched the horizontal agent pitch for two years. It always sounds the same. "Give the model a computer, give it tools, and let it figure out what to do." The pitch deck has the same screenshots: an agent browsing a website, clicking a button, sending an email, booking travel, answering a support ticket — all in one continuous task. The demos are always impressive.
What the pitch deck does not have is a paying customer who renewed after the second invoice. I have watched three horizontal agent startups hit the same wall in twelve months. The horizontal agent fails on the long tail of tasks that domain experts handle in their sleep. The booking flow breaks on a third-party calendar integration. The travel agent books a hotel that closed eight months ago. The support agent promises a refund the system cannot issue. The cost of failure is unbounded. The blast radius is uncapped. The customer churn is terminal.
The model is a tool inside a deterministic system, not the spine of one. The horizontal pitch gets that exactly backwards.
Look at the agent products that crossed $50M in annual revenue in 2025 and 2026. Harvey, the legal AI. Devin and Cursor, the agentic coding surfaces. Glean, the enterprise search agent. Hippocratic AI, the clinical workflow agent. Sierra and Decagon, the vertical customer-service agents. These are not horizontal agents. They are domain-specific, data-rich, deeply integrated, and narrow on purpose. They were built by people who understood the workflow before they ever wrote a system prompt.
Harvey does not book travel. Devin does not answer HR questions. Glean does not design proteins. The narrowness is the product. The vertical agent wins because it can be tuned, evaluated, and trusted inside a domain its builders understand. The cost of failure is bounded. The customers renew because the agent does the one thing they hired it to do, well enough that human reviewers trust it.
The horizontal pitch is irresistible because it is the pitch venture capital wants to fund. A vertical agent caps at a TAM you can size. A horizontal agent is "the agent for everything" with a TAM that ends in "trillion." The pitch gets a Series A. Two years later the team has an agent that does twenty things at 80% accuracy and gets replaced by a domain-specific tool in every category that matters. The thesis was wrong.
The model labs keep shipping demos that look horizontal — Claude with computer use, OpenAI's Operator, Google's Astra. The demos do not survive contact with a paying customer who has a regulated workflow. The labs ship them anyway because demos drive API consumption, and API consumption is the metric they report to the board. Your startup is not the model lab. Build the thing that survives the second invoice.
If you are building a horizontal AI agent right now, you are building a 2024 demo. Stop. Pick a vertical. Pick a workflow. Pick a customer segment. Build the deepest possible agent for that segment and own it. The horizontal agent is what gets a TechCrunch headline and a non-renewal. The vertical agent is what gets a multi-year contract and a case study. The teams that listen now will own the next decade of agentic AI. The teams that do not will be writing postmortems about why their "agent for everything" got out-executed by a Harvey, a Glean, or a Hippocratic in every category that mattered.
— *Mr. Technology*
*Posted June 8, 2026. The horizontal agent is a demo. The vertical agent is a business. Pick a side.*