The Code Behind AI: Why Software Trumps Language Models

Language Models">

The development of autonomous AI agents is often touted as a revolution driven by advancements in language models. However, a new review paper argues that the real bottleneck for these agents isn't the language model itself, but rather the software layer wrapped around it. This layer, which includes tools, memory, testing, and permission boundaries, is what transforms a stateless model into a working agent.

The paper's thesis is already being put into practice by Deepseek, a company that is building a dedicated "Harness" team in Beijing. According to Deepseek, the core formula for creating AI agents is simple: model plus harness equals AI agent. This approach highlights the critical role that software plays in enabling language models to think and act autonomously.

The distinction between the language model and the software layer is crucial, as it challenges the common perception that AI agents are solely driven by their language models. In reality, the software layer is what enables these models to interact with their environment, make decisions, and take actions. By focusing on the software layer, researchers and developers can create more sophisticated AI agents that are capable of complex tasks.

As the field continues to evolve, it will be interesting to see how the interplay between language models and software layers shapes the future of autonomous AI. The new review paper provides a timely reminder that the development of AI agents is not just about improving language models, but also about creating the software infrastructure that allows them to operate effectively. The article's findings have significant implications for the future of AI development, and companies like Deepseek are already taking steps to harness the power of software to create more advanced AI agents.