Why We Will Not Invest in Horizontal AI Agents

The most crowded category in artificial intelligence today is also, in our view, the most likely to be commoditized within the next twenty-four months. Horizontal AI agents — broad, general-purpose autonomous assistants that promise to handle any task across any function — have attracted a disproportionate share of venture capital, founder ambition, and public attention. We believe this attention is misplaced. New Future GPT LLC will not allocate capital to this category, and the reasons behind that decision sit at the center of how we think about durable value in the AI economy.

The structural problem with horizontal

A horizontal AI agent, by definition, must do many things adequately rather than one thing exceptionally. This is a serviceable design philosophy for a research demonstration. It is a poor design philosophy for a defensible business. The companies building horizontal agents are competing in a market where the underlying capability — general reasoning, tool use, and task decomposition — is being commoditized by the same frontier laboratories that originally invented it. Every six months, the baseline performance of a general-purpose agent improves, free of charge, because OpenAI, Anthropic, Google, and a handful of others release new model versions. The startups building thin wrappers on top of those models are not capturing the value. They are renting it.

This dynamic is not new. It is the same dynamic that destroyed dozens of horizontal chatbot companies between 2023 and 2025. The companies that survived were either the frontier laboratories themselves, or those that built deeply specialized vertical products that the laboratories could not easily replicate. The horizontal category, in the chatbot wave, was a graveyard. The agent wave will follow the same pattern.

What we believe will survive

Vertical agents will survive. We define a vertical agent as one that is built for a single industry, embedded in that industry’s systems of record, equipped with that industry’s regulatory and operational context, and sold through that industry’s procurement processes. A legal research agent that understands the difference between case law in Delaware and California is a vertical agent. A claims adjudication agent that knows the difference between a homeowner’s policy and a commercial auto policy is a vertical agent. A clinical documentation agent that operates inside the electronic health record of a specific hospital system is a vertical agent.

These products share four characteristics that horizontal agents lack. They own a workflow rather than suggesting one. They integrate with proprietary systems that cannot be replicated by a frontier laboratory. They are evaluated on outcomes that matter to the buyer — billable hours saved, claims processed, charts completed — rather than on demonstration metrics. And they are sold to people who buy software for a living, not to consumers who download applications on impulse.

The unit economics also differ. A horizontal agent must compete on price because its underlying capability is increasingly free. A vertical agent competes on value because no frontier laboratory will ever build the same product. One business has gross margins that erode quarterly. The other has gross margins that expand as the product becomes more deeply embedded in customer operations.

The exception that proves the rule

There is one category of horizontal agent that we believe will produce extraordinary returns: the frontier laboratories themselves. OpenAI, Anthropic, Google, and a small handful of others will build the most capable general-purpose agents because they own the underlying models. These companies do not need venture capital from a holding firm like ours. They are competing for sovereign-scale infrastructure investments, not Series A rounds.

For every other participant in the horizontal agent market, the question is not whether they can build something useful. They probably can. The question is whether they can build something defensible. The historical evidence on thin wrappers over frontier models is clear, and the future evidence is unlikely to be different.

What this means for our portfolio

We will continue to evaluate companies in the vertical agent space — for legal, financial, healthcare, industrial, and procurement applications. We will continue to invest in the operators who understand that an agent in production is fundamentally different from an agent in a demo. And we will continue to decline meetings with founders who answer the question “what is your moat” by referencing the size of their seed round.

The frontier laboratories will build horizontal intelligence. The defensible companies will build vertical work.