October 13, 2026Main stage3 min read

Hermes gives you wings: Build on open foundations.

Dillon Rolnick makes the case for owning the agent layer as AI becomes part of everyday business operations. His argument connects model economics to a practical architecture: interchangeable models beneath software and workflows a company controls.

Dillon RolnickNous ResearchCEO

Watch "Hermes gives you wings: Build on open foundations. — conference livestream" on YouTube

From the conference livestream · 7:21:20–7:36:24 on YouTube. The player starts and stops at this session’s boundaries.

7:21:21 ↗

When switching models is easy, where does the business settle?

Practitioners switch models readily. A new release appears, they try it, and their preference can change again a week later. Dillon Rolnick opens by asking how AI companies earn back large infrastructure and research investments if the intelligence they sell becomes increasingly interchangeable.

His answer is an argument about incentives. Companies can move toward infrastructure or closer to the user, where integration into daily work can make a service harder to replace. Or they can pursue sufficiently powerful systems to solve valuable problems themselves. These are competing strategies in Rolnick's account of the market.

The user-facing opportunity leads him to business agents. Businesses have economically valuable uses for tokens and workflows into which software can become deeply embedded. The concern is what happens as more of a firm’s work depends on an agent platform controlled by someone else.

Rolnick’s forecast is deliberately provocative: if a large share of future output comes through that platform, its operator can gain considerable negotiating power. A change in price, access, or terms can then affect far more than one employee’s chatbot subscription.

7:25:02 ↗

An agent accumulates the company’s way of working

The other possibility Rolnick raises is direct competition. He believes model developers may eventually apply their systems to lucrative markets themselves, with finance an attractive candidate. He describes today’s partnerships and hiring as a way to learn the context they currently lack. He is forecasting how those incentives may play out, rather than describing an announced plan by a particular firm.

For the audience, the immediate issue is where to build. Tools acquire an organization’s know-how over time. People write processes around them; systems connect to them; working habits settle. Moving later can mean changing a substantial part of how the business operates.

Rolnick’s proposal is to put that accumulating work on open foundations, specifically Hermes Agent. He introduces Nous Research’s preference for decentralization as a philosophical commitment, while also arguing that the business economics can align with it.

His science-fiction joke—“Skynet was, after all, a closed source company”—captures the tone. The practical claim underneath it is that control over a widely used layer of intelligence carries power, and the organization using that layer should retain an exit.

7:29:16 ↗

Software features are easier to copy than trust is to earn

Rolnick argues that the cost of reproducing software has fallen enough to weaken proprietary features as a lasting defense. He gives a company example: the team liked an interface released by a competitor, used Hermes Agent to reproduce the concept over a weekend, and offered the resulting feature as open source.

From that example he draws a broader bet on network effects, reputation, and reliability. If open-source builders can reproduce useful functionality quickly, keeping the implementation private buys a vendor less time.

He makes a related argument about models teaching other models. In his account, much of the performance advantage lies in data quality and scale, and distillation allows some capabilities to transfer. He criticizes the language used to describe unauthorized distillation and restrictions on using models to train other models. His criticism concerns industry incentives; whether a particular training arrangement is lawful is a separate question.

Fast-changing features and models make substitutions likely. An agent architecture can let a business change those parts without replacing its accumulated workflows.

7:32:25 ↗

Build an application that combines rules and judgment

Rolnick describes agents as closer to operating systems and frameworks than to chatbots. He explicitly returns to the earlier investment-firm presentation: data sources, ordinary code, people’s opinions, and judgment combine into a decision-making application.

Deterministic and nondeterministic work can coexist. Some steps should execute rules; others need reasoning or interpretation. A business can compose increasingly sophisticated workflows from those pieces while retaining the ability to change the model underneath them.

Owning the foundation does not mean becoming a full-time agent-framework company. Rolnick argues that teams should focus on their own work and inherit improvements from a shared system as the state of the art changes. Hermes is his proposed base: run it where needed, modify it, and retain the ability to leave.

He closes with the business he wants to build around that arrangement. A widely adopted open standard creates demand for support, safety, scalability, and reliability. Nous can compete for that work through reputation rather than making exit impossible.

For builders, the proposed bargain is straightforward: concentrate on the valuable work you understand, inherit improvements from a shared agent framework, and retain the ability to change its model or operator. Nous would have to keep earning the support business even when customers are free to leave.

When switching models is easy, where does the business settle?

An agent accumulates the company’s way of working

Software features are easier to copy than trust is to earn

Build an application that combines rules and judgment

Further reading