Dillon Rolnick on AI Vendor Lock-In and Open Agent Foundations

I am Dillon Rolnick, CEO of Nous Research. I want to explain why I believe companies should build AI systems on foundations they can control, and how my company hopes to support that approach.

AI labs have made enormous investments in research and infrastructure. At the same time, I believe models are becoming more interchangeable: practitioners switch among providers, and token prices are declining even as usage grows.

My interpretation is that competition compresses the margins available from selling model intelligence alone. These are my market judgments; the talk does not establish a complete accounting of industry investment or profitability.

I see two possible responses from labs trying to earn back their investment: compete elsewhere in the stack, closer to infrastructure or end users, or develop much stronger intelligence and use it to pursue valuable problems directly.

The agent layer matters in this account because software that becomes embedded in a business can be difficult to replace. I expect vendors to pursue enterprise customers whose AI use has clear economic value.

I speculate that agents could account for a large share of business output in a few years. If another company controls those systems and its incentives diverge from yours, dependency could give it leverage over your operations; this is a warning, not a demonstrated forecast.

I also suggest that a provider with very capable AI might choose to solve lucrative problems itself instead of only selling access to a model. In my finance example, I imagine firms learning from partners before potentially competing more directly.

Vendor lock-in is not a new concern, but I think the stakes rise when AI becomes embedded in many decisions and processes. Organizations choosing tools now will accumulate skills and habits around them, which makes the choice of foundation consequential.

My conclusion is that businesses should favor open foundations rather than build on what I call quicksand. I advocate Hermes Agent, a product from my own company, so this recommendation is also a commercial position.

We founded Nous Research because we were concerned about concentrated control over an important technology. I frame that as both a philosophical commitment and a business model in which our interests can align with users.

I argue that reproducing software has become cheaper. As an example, I say our team used Hermes Agent to reproduce an interface resembling Grok bots over a weekend and released its implementation openly; I use that anecdote to argue that proprietary features may be copied faster.

In my view, durable advantages will shift toward network effects, reputation and reliability. That claim is a prediction about competitive dynamics, not proof that every proprietary software moat has disappeared.

I then discuss distillation: using one model to help train another. I criticize how some model companies characterize or restrict the practice, but the legal and contractual status of particular methods is disputed and is not resolved by this talk.

I see agents as closer to operating systems or frameworks than simple chatbots. An agentic application can combine deterministic code, data sources, documents, human opinions and model-generated reasoning or judgment.

Companies should be able to swap models while building functionality on top of those primitives. They should also keep their attention on their own strengths rather than spend all their resources becoming intelligence-model companies.

As agent systems evolve, I want users to inherit improvements without rebuilding everything. I present Hermes Agent as a system that can be run in different places and modified, though adopting any tool still requires assessing its actual license, operations and dependencies.

I want Hermes Agent to become an open standard and claim it is widely used relative to competitors. Those adoption claims are my company’s statements in the presentation, not independently verified figures here.

I close by urging builders to retain autonomy as they use agentic systems to solve difficult and valuable problems. My proposed path is to choose a foundation that users can operate and change themselves.