AI Loops: Now agents can run jobs on their own.

Boris Cherny, the engineer behind Claude Code, demonstrated a new type of AI agent at the recent @Scale conference. His agents are not just single agents that operate on your behalf, but rather loops of many background agents that operate without needing to be told what to do. These agents roam around in your codebase, spotting redundancies and other problems, fixing them, and even automatically opening pull requests (without any human intervention!) The entire process is completely transparent to him; he doesn’t know what tasks his agents have chosen, nor does he need to approve their actions.

This differs from most of the existing examples of agentic AI. In those cases, the interaction is generally transactional, where you specify a goal, the agent completes some work towards that goal, you inspect the work, and then you specify another goal. With loops, there is no need for this back-and-forth; instead, the agent observes the world, determines what needs to be done, completes the task, and repeats. It’s like having a background process, except it isn’t responding to a prompt.

1The 'Ralph Loop' — a practical technique already in use

An interesting side effect of these loops is that they help keep models focused on their goals over the course of a longer session. To do this, many developers have created something called the Ralph Loop, which involves the agent summarizing its actions after completing each task, and then asking itself whether it has accomplished its overall goal. This helps ensure that models don’t get distracted while they are working, and can also help detect cases where an agent has been completing tasks that satisfy the formal requirements, but aren’t actually helpful.

2What this means for how you'll work

In general, the philosophy of loops fits in with the general trend towards thinking about problems iteratively, using o3-style reasoning models, and long agent runs. But with loops, that iteration happens over time, without you being present.

Of course, there are also some downsides. For one thing, loops require continual token expenditure, and if you have a horde of background agents operating without any constraints, you may end up getting stuck with an unexpectedly large bill. But that shouldn’t discourage you; instead, just make sure you have a clear cutoff point and budget when you get started. Developers who figure out how to operate loops effectively will have a significant advantage once these features are available in products like Claude Code.