At Meta’s @Scale AI & Data 2026 conference, Claude Code head Boris Cherny described “loops” as a new abstraction in AI-assisted software development: agents prompting other agents to write and improve code. TechCrunch reported that Cherny rejected the idea that the pattern is merely hype, while Dealroom.co summarize...
Anthropic’s Boris Cherny used a Meta @Scale conference fireside chat to argue that AI-assisted coding is moving from one-off prompts toward continuously running “loops.”
The official @Scale Conferences agenda for AI & Data 2026 lists a session titled “Fireside Chat with Boris Cherny, Head of Claude Code,” scheduled for June 22, 2026, with Cherny as speaker and Meta’s Jesse Chen as moderator. The @Scale video page also identifies the session as a fireside chat with Cherny of Anthropic and Chen of Meta.
The discussion centered on how software engineers are using AI coding systems and what comes after the current wave of agent-assisted development. According to TechCrunch, an audience member asked Cherny whether AI “loops” were just another hype cycle.
Cherny’s answer, as reported by TechCrunch, was that the pattern is real and important: agents can now prompt other agents that write code. Dealroom.co similarly summarized Cherny’s framing as a progression from engineers writing source code by hand, to agents writing code, to agents prompting agents that write code.
In this context, a loop is not a new programming primitive. It is a workflow in which an AI system repeatedly performs a task, evaluates progress, and decides what to do next, often with limited direct human prompting.
Dealroom.co reported that Cherny described loops as the next abstraction after the move from source code to agents. TechCrunch characterized the idea as part of a broader “loopy” turn in AI development, where engineers set up systems that keep working through tasks rather than waiting for a single prompt-and-response exchange.
That distinction matters for software teams. A prompt-based workflow usually begins with a human asking an AI system to write, refactor, or explain code. A loop-based workflow can be designed to keep checking a codebase, proposing changes, or pushing work forward until a stopping condition is reached.
Cherny’s comments do not mean that human developers disappear from the process. The sources describe a shift in where the developer’s effort goes: less time spent issuing individual instructions, and more time spent designing the conditions, objectives, and safeguards under which AI systems operate.
The interest around loops reflects a practical question now facing engineering teams using AI coding tools: how much work should be delegated to systems that can keep iterating?
TechCrunch reported that Cherny rejected the suggestion that loops are only hype. Dealroom.co reported that he sees loops as a major step in the abstraction stack for programming. Taken together, the accounts point to a debate that is less about whether AI can generate code, and more about how software work should be organized when AI tools can repeatedly plan, write, and revise.
The @Scale setting is also notable. Meta’s @Scale conferences are aimed at technical audiences working on large-scale systems, and the AI & Data 2026 agenda placed Cherny’s session among broader discussions of AI infrastructure and data systems. That context helps explain why the conversation focused not just on model capability, but on the engineering patterns emerging around those models.
The sources do not establish that loops are a settled best practice, nor do they show that every software team should adopt them. The reporting instead shows that prominent AI coding practitioners are beginning to name and discuss a pattern already visible in advanced AI-assisted development.
The unresolved issues are significant. Continuously running AI workflows can consume more compute, create more code for humans to review, and introduce new risks if their goals are poorly specified. They may also require stronger evaluation systems, clearer permissions, and better cost controls than ordinary prompt-based tools.
Cherny’s @Scale remarks, as reported by TechCrunch and Dealroom.co, suggest that AI coding is entering a phase where the main skill is not only asking a model for code. It is designing repeatable systems that decide when to ask, what to change, how to check the result, and when to stop.
The @Scale video page also identifies the session as a fireside chat with Cherny of Anthropic and Chen of Meta.
The discussion centered on how software engineers are using AI coding systems and what comes after the current wave of agent assisted development.
According to TechCrunch, an audience member asked Cherny whether AI “loops” were just another hype cycle.
Continue reading
