AI models are becoming interchangeable, so labs will try to recover their investments by controlling the agent layer or competing directly in profitable industries.
2
Companies that build critical workflows on closed AI platforms may become dependent on vendors whose incentives conflict with theirs.
3
Businesses should build agent systems on open foundations they control, so they can change models and inherit new capabilities without rebuilding their applications.
Summary
Dillon Rolnick argues that falling model prices will force AI labs to find other ways to recover the enormous cost of training and serving their systems. He expects labs to embed themselves in businesses through agents, then raise prices, or use their own systems to compete directly in lucrative industries such as finance. That makes the choice of technical foundation a business decision. Rolnick says proprietary software has lost much of its old protection because agents can reverse-engineer features quickly. He describes agents as operating-system-like systems that combine code, data, deterministic workflows, judgment, and model-based reasoning. Companies should own the foundation beneath their applications, keep the ability to swap models, and let improvements in agent technology flow into their products. He presents Hermes Agent, Nous Research's open-source project, as an example of this approach and argues that future moats will come from reputation, reliability, and network effects.
Rolnick says AI labs have spent hundreds of billions of dollars individually, while the industry may have spent trillions. At the same time, models are becoming interchangeable for practitioners, who switch among providers to try new systems. He points to token prices falling while total token use rises and infrastructure costs remain flat. That combination narrows the margin available from selling model access. Labs therefore need another layer of the stack where they can capture more value.
Labs can recover their spending through control or direct competition
Rolnick gives two paths for labs trying to make back their investments. They can move down into infrastructure or up toward users, where they can create stickiness and protect margins. His other path is to build extremely capable systems and aim them at valuable problems themselves. He says the fight has moved toward agents because agents can sit inside a user's work rather than being an interchangeable model endpoint.
Businesses are the target because their AI use is tied to revenue
Consumers are fickle customers, Rolnick says, while businesses spend heavily on tokens for economically productive work. Labs therefore have an incentive to become embedded in companies. He imagines a future where much of a workforce's output is organized by agents, creating dependence on whoever controls those systems. A vendor with that position could use its leverage to change the rules or raise prices.
Finance is an obvious industry for labs to attack directly
Rolnick says labs may eventually point their systems at the most lucrative problems instead of selling model access at falling margins. He describes financial markets as problems the labs believe are solvable and relatively tractable. They currently lack enough industry context, experience, and know-how, so they are hiring investment banking analysts and partnering with large firms. His expectation is that they will eventually move closer to the source of the money.
Companies should avoid foundations they cannot control
Rolnick warns that the tools and processes companies choose now will become embedded in their organizations. Teams will build know-how, workflows, and systems around those choices, making them harder to change later. He calls closed dependencies 'quicksand foundations' and urges companies to use open foundations instead. His concern is that a business could become dependent on a platform whose incentives are not aligned with its own.
Open-source builders can reproduce proprietary software quickly
Rolnick says proprietary software no longer has the protection it once had because the cost of reproducing software has fallen. He gives Grok bots as an example: Nous Research pointed Hermes Agent at the interface, reproduced it over a weekend, and released the result as an open-source feature. He argues that motivated open-source developers can use agents to reverse-engineer interesting features soon after they appear. Future advantages, in his view, will come from network effects, reputation, and reliability.
Distillation makes model advantages easier to copy
Rolnick says models can teach other models to become nearly as capable. He argues that model advantages usually come from data quality and scale rather than from a proprietary architecture. Model companies call the teaching process distillation and may call unauthorized distillation theft. He also criticizes the practice of training models not to help with the creation of other models, which companies describe as safety.
Agents combine ordinary software with judgment and reasoning
Rolnick asks developers to think of agents as operating systems or frameworks rather than personal assistants. An agent can combine data sources, routine code, documents, human opinions, and judgment. The resulting application does not need to be entirely deterministic or made only from fixed code. It can add reasoning and intelligence to existing systems while still relying on ordinary software components.
Owned agent foundations let businesses inherit improvements
Companies should build their applications on foundations they own and should be able to swap models in and out. Rolnick says they should create layers of functionality on top of primitives such as code, data sources, documents, reasoning, and intellect. Most companies do not need to become AI or agent companies themselves. They should inherit improvements in agent systems instead of rebuilding and maintaining those systems whenever the state of the art changes. He presents Hermes Agent as an open foundation that can run wherever its users choose and be modified by them.
"If you're going to be competing at the user layer, and you're going to be bringing this fight somewhere else, you're going to be bringing it to businesses."02:42
Who should watch
You are deciding whether your company should build important workflows on a model provider's agent platform.
Your team needs a way to combine models with existing code, data, documents, and human judgment without becoming dependent on one vendor.
You want to understand why an open-source agent foundation could matter even when the underlying models keep changing.