The shift toward continuous agentic loops in AI

On Friday, Boris Cherny, the creator of Claude Code, spoke at Meta’s @Scale conference about the emergence of agentic loops, which are recursive processes where AI agents work continuously. When asked by an audience member if loops are simply the next hype cycle or if they are real, Cherny asserted that they are real. He noted that two years ago, developers wrote source code by hand, before transitioning to a phase where AI agents write the code. Now, Cherny explained, the industry is transitioning to a point where agents prompt other agents to write the code. According to Cherny, “As big as the step from source code to agents was, loops are just as important and as big a step.”

In his own workflow, Cherny keeps specific loops running continuously. For example, around the 32:00 mark of his presentation, he described using one agent to constantly look for ways to improve code architecture, while another searches for duplicated abstractions to unify. These agents submit pull requests like any other coder, running endlessly as the code changes. To manage these continuous processes, developers use techniques like the Ralph Loop, where a model sums up its work to check if a goal is accomplished. This continuous operation aligns with a broader industry push toward “test-time compute,” which refers to additional compute resources used during inference to improve model performance. OpenAI researcher Noam Brown recently observed that contemporary AI models can solve nearly any problem if they are provided with sufficient compute. Agentic loops operationalize this by continuously applying compute to tasks, such as code base improvements, until a goal is met.

However, this continuous operation introduces financial trade-offs. Because agentic loops run endlessly in the background, they consume tokens at a high rate with no natural ceiling on spending. While this high consumption benefits companies like Anthropic, which operate on a token-selling business model—charging users based on AI model usage—it presents a challenge for other enterprises. For users outside of this business model, running continuous agentic loops may be a costly way to work.

Why it matters

The shift toward agentic loops, where AI agents work continuously in the background, represents a potential next step in AI capability, though it introduces significant costs and risks regarding token consumption and model drift.