Agentic AI Loops: The Next Leap in Automation

Boris Cherny, creator of Claude Code, recently spoke at Meta's @Scale conference, where he fielded a surprising question from the audience: are loops the future of AI development? Cherny's response was unequivocal: loops are a game-changer. He explained that his team has transitioned from writing code by hand to using agents that write code, and now they're taking it a step further by having agents prompt other agents to write code. This approach, known as agentic loops, allows for continuous improvement and automation.

Cherny shared examples of how he uses loops in his own work, including having one agent optimize code architecture while another identifies duplicated abstractions. These agents submit pull requests like human coders and run indefinitely, making them a powerful tool for streamlining development. While the concept of loops is not new, the application of agentic loops in AI development is a significant shift. It requires a high degree of trust in AI, but with models improving rapidly, it could be the key to unlocking real productivity gains.

However, this approach also raises concerns about cost, as agentic loops can be computationally intensive and expensive. Despite these challenges, the potential benefits of agentic loops make them an exciting development in the field of AI infrastructure and automation.

Source: TechCrunch