Claude Code: Agentic Loops Revolutionize AI

Claude Code and the Impact of Agentic Loops

At the @Scale conference organized by Meta, Boris Cherny, the creator of Claude Code, addressed a question from the audience regarding loops, which are often perceived as a passing trend. Cherny quickly dispelled this doubt by confidently asserting that loops are indeed real and not just a fleeting fad. His emphatic response to the question was, “Yes, they are real.”

Cherny explained that the development of AI has undergone several significant phases. Two years ago, the source code was still predominantly written by hand. Then, a transition occurred towards the use of agents capable of generating code. Today, we are reaching a stage where these agents can prompt other agents to produce code, marking an advancement as significant as the initial shift from source code to agents.

Autonomous Agents in Action

In his talk, Cherny detailed how he applies loops in his own work, specifying these aspects around the 32:00 mark in the YouTube video of his presentation. One agent is constantly seeking ways to optimize the architecture of the code, while another focuses on detecting and unifying redundant abstractions. These agents submit pull requests just like a human developer would, and their activity is continuous, as the code evolves incessantly.

This approach, supported by a figure as influential as Cherny, highlights the potential of loops in AI. With the rise of autonomous agents, the goal is to manage these entities effectively: setting clear objectives, tracking measurable progress, and preventing them from straying too far from their initial mission. Loops go further by allowing a network of agents to operate continuously, a perspective that relies on increased trust in AI. That’s a lot of trust to place in AI — but with rapidly improving models, this could be the next step in enabling AI to handle real work.

Revisited Recursive Loops

Although the idea of loops may seem new, it is rooted in a well-established tradition in computer science. Recursive loops, which involve functions calling themselves to accomplish a repetitive task, are a fundamental concept taught in early computer science courses. In the context of agentic loops, this logic is applied to agents that decide for themselves when to stop the loop, introducing a non-deterministic dimension — that is, a sub-agent chooses when to stop the loop instead of a clear condition.

One of the simplest and most popular loops is the Ralph Loop. It involves summarizing the work done by the model and checking if the goal has been achieved. This method helps manage AI models that might get lost when operating for too long, bringing them back to their starting point until the mission is accomplished.

Towards Intensive Computation During Testing

Loops also fit into a broader trend towards intensive computation during the testing phase. Noam Brown, a researcher at OpenAI, recently emphasized that current models can solve almost any problem, provided there is enough computing power. This approach is particularly relevant for hill-climbing type problems, where the model continues to make incremental improvements until reaching a predefined threshold.

However, this method is not without cost. Agentic loops, like AI agents, consume tokens at a much faster rate than simple chatbots. Maintaining a loop in continuous operation entails potentially unlimited expenses, which can be costly for businesses. This suits Anthropic, which is ultimately in the token-selling business.

An Expensive but Potentially Profitable Investment

Despite the associated costs, agentic loops could prove to be a wise investment. Depending on the problem to be solved and the appropriate setup to monitor token expenditures and control drift, the benefits can far outweigh the expenses. This technology promises to transform the way AI is used to manage complex tasks, offering increased autonomy and efficiency.