According to the external review, AI's development community has recently begun to focus on a modus operandi known as “cycle”: Instead of simply allowing individual agents to perform one mission, multiple agents would continue to operate in the back office and work with each other to improve the same job over and over again. This may move the AI programming tool from a single response to a continuous mission.
Meta talks about "cycle."
At the @Scale conference in Meta, Claude Code founder Boris Cherny was asked whether the "circle" was a new round of open-up or was of real value. According to him, this direction has already entered the real-life phase.
Cherny states that the software development process is changing continuously. First the developer's hand-written code, then the AI proxy generation code, now becomes the proxy pointer, and then the latter completes the specific writing and modification. According to the article, in his view, this step was no less important than the previous shift from handwritten code to AI agent.
Multiple proxy continuous code change
Cherny cited, for example, one agent, who was constantly looking for ways to improve the code structure, and the other, who was specializing in replicating abstractions and trying to merge. They will be submitted like ordinary developers. As the coding library itself is constantly changing, these agents will not stop running.
According to the article, this model differs from the way most current users use AI agents. It was more common in the past to set clear targets for agents, to check results by stage and to limit their deviations from the hints to the extent possible. The so-called “cycle” is to further liberalize the mandate and allow a group of agents to carry out their duties in the back office for a long period of time.
More potential, more cost.
TechCrunch notes that this is not an entirely new concept of computer science. Recursive loops are the basic method of programming, except that the current conditions for cessation are no longer determined solely by the fixed rules but may be left to another AI sub-agent. That is to say, the idea of AI monitoring AI is moving from theoretical structure to actual product form.
The article also links this trend to “test-time calculus”. OpenAI researcher Noam Brown mentioned earlier this month that the current model has been able to deal with a large number of issues with sufficient input. Under this logic, if a task can be gradually improved over time, for example, by optimizing the coding library, continuous investment in computing may result in continuous change of results.
However, the cost of this approach is also direct. According to the article, the proxy AI would have consumed more token than the regular question-and-answer chat robot, and the cost cap would have been more difficult to control because the circulation of the work stream was operated by default. Long-term operation of multiple agents may be an expensive way of working for users.
Despite this, it is argued that there may still be significant gains from such revolving work flows as long as the task itself is adequate for continuity and as long as the enterprise is able to monitor token expenditures, mission deviations, etc.