Investing in Autonomy: Ten Hypotheses

Variant invests in technology that expands autonomy — a person or organization’s freedom to build, customize, and act on their own terms. 

Automation helps developers, individuals, and organizations do more with less. Done right, it increases agency for those at the helm by giving the user both new capabilities and greater ownership over their stack. 

Such surface area is (rightfully) quite broad. So, as we continue to explore this emerging domain, we’ve found it useful to anchor our thinking in a small set of hypotheses about how the world may evolve.

The following are ten working hypotheses that have emerged from our discussions over the past few months. They’ve become a useful lens for deciding where we spend our time and attention. Understandably, some of these hypotheses may prove incomplete (or simply wrong). But we’ve found this exercise useful because it forces us to make our assumptions explicit. It gives us a framework for thinking about how the world may change, what products will matter, where value may accrue, and who is likely to capture it.

1. Agents will become the dominant source of traffic on the internet. Much of today’s software assumes a human is clicking buttons and reading interfaces. Increasingly, products will need to be designed for both humans and agents. Agents operate continuously, are infinitely replicable, and have fundamentally different behaviors and incentives than people.
2. The world will remain compute-constrained for the foreseeable future. Scarcity makes infrastructure matter. Startups that reduce the cost of compute, improve utilization, or unlock new sources of supply become increasingly valuable.
3. Open source will play an increasingly important role across the AI stack. For enterprises, open source provides a credible strategy for competing with the frontier labs. For developers and consumers, it lowers the cost of experimentation and expands access to state-of-the-art models.
4. Access to frontier AI will become increasingly permissioned. Demand will grow for open, permissionless alternatives that preserve developer autonomy.
5. There will be an order of magnitude more models that matter. As models become more specialized, selecting and orchestrating them becomes more valuable than building around any single model.
6. Multi-model systems will become the default. Applications, not users, will manage the complexity of routing work across many models.
7. Harnesses are a major source of value accrual within the AI stack. As models become easier to substitute, the durable relationship shifts to the product that orchestrates models and owns the user experience.
8. Many agentic systems and tools today are single-player; the frontier is multiplayer. The hard problem shifts from building individual agents to coordinating shared context, permissions, workflows, and collaboration across organizations. Defining and owning the control plane is a major point of leverage.
9. AI will create lower-cost alternatives to many of society’s most expensive services. Industries such as healthcare, education, and legal services are especially ripe for new entrants that rethink the economics of delivering expertise. Some of the most novel and surprising new solutions may stem from empowering the end user with new capabilities.
10. AI will make many previously inaccessible markets economically viable for startups. Entrepreneurship becomes more tractable in more domains as agents automate more of the work required to start and scale a company, allowing founders to focus on the hardest problems.

We’re putting this out as a public scorecard of our hypotheses as of early Q3 2026. We’re excited to track how it holds over time, and are looking forward to updating our beliefs as they evolve.

We’re also actively investing across these hypotheses. If you’re building against one of these hypotheses, we’d love to talk. (And especially if you disagree, please tell us where we’re wrong!)

Thank you to Jesse Walden, Jay Drain Jr, Caleb Shack, and the Variant team for thoughts and feedback that contributed to this piece.