Why AI Agents Should Be Built on Code, According to Lenny Rachitsky

In a recent LinkedIn post, Lenny Rachitsky highlights insights from Alexander Embiricos, Product Lead at OpenAI, regarding the future of AI agents and their interaction with computers. Rachitsky shares Embiricos’s perspective that the most effective path for AI models to perform tasks involves writing code, rather than relying on less predictable methods like OS hacking or point-and-click interfaces.

The Case for Coding Agents

Rachitsky, relaying Embiricos’s argument, posits that as AI development progresses, the creation of AI agents focused on coding might become a primary objective. This is because for AI to function as a “super assistant” capable of executing actions, it needs a robust mechanism to interact with and control a computer. Embiricos suggests that traditional methods of AI-computer interaction are suboptimal.

“You could try to hack the OS and user accessibility APIs. Maybe you could point and click. That’s a little slow and unpredictable.”

As Rachitsky conveys Embiricos’s view, these methods are fraught with inefficiencies and lack the reliability needed for sophisticated AI operations. The inherent unpredictability and slowness of such interfaces limit the potential of AI assistants.

Code as the Optimal Interface

The core of Embiricos’s argument, as shared by Rachitsky, is that the most effective and efficient way for AI models to leverage computing power is through the generation of code. This approach offers a direct, powerful, and predictable means for AI to interact with digital environments and accomplish tasks.

“Another way, and it turns out the best way for models to use computers, is simply to write code.”

Rachitsky emphasizes that this perspective suggests a significant shift in how we should conceptualize and build advanced AI. Instead of trying to make AI mimic human-computer interaction through graphical interfaces, the focus should be on enabling AI to operate at a more fundamental level by producing code.

Implications for Future AI Development

The insights shared by Rachitsky point towards a future where AI agents are not just passive recipients of instructions but active creators and manipulators of digital systems through code. Embiricos, according to Rachitsky, believes that this capability is fundamental for building truly capable AI.

“We are getting to the point where if you want to build any AI agent, maybe you should be building a coding agent.”

This viewpoint, as highlighted by Rachitsky, suggests that AI agents built with the primary function of writing code may unlock unprecedented levels of functionality and utility. It implies that the development of more sophisticated AI assistants hinges on their ability to programmatically interact with and control their environment, positioning coding agents as a critical frontier in AI research and development.