Claude Code's Boris Cherny says AI 'loops' are the next real shift in software development
At Meta's @Scale conference, the creator of Claude Code argued that continuously running agent swarms represent a leap as significant as the move from hand-written code to AI agents.
What matters
- Claude Code creator Boris Cherny told Meta's @Scale conference that AI 'loops' are real and represent a shift as big as the move from hand-written code to AI agents.
- Loops authorize swarms of agents to run continuously in the background, finding and submitting improvements without human initiation for each task.
- Cherny described his own setup: one agent perpetually improves code architecture, another unifies duplicated abstractions, both submitting pull requests like human developers.
- The model inverts current agentic AI practices, where humans carefully manage each agent invocation and check in on discrete progress.
- Key open questions include review burden, cost of continuous compute, and code quality oversight.
What happened
On Friday, Boris Cherny — the creator of Anthropic's Claude Code — appeared at Meta's @Scale conference and fielded an unusually pointed opening question from the audience: "Are loops the next hype cycle, or are they for real?"
Cherny's answer was unambiguous. "Yes, they're for real," he said.
He framed loops as the latest in a series of major transitions in how software gets written. "Two years ago, we wrote source code by hand. We started to transition so agents write the code. And now we're transitioning to the point where agents are prompting agents that then write the code," Cherny explained. "As big as the step from source code to agents was, loops are just as important and as big a step."
Later in the talk (around the 32:00 mark in the conference video), Cherny got concrete about the loops he runs in his own work. One agent continuously scans for opportunities to improve code architecture. Another looks for duplicated abstractions that can be unified. Both submit pull requests just like any human contributor — and because the codebase is constantly changing, the agents never stop running.
The concept is straightforward but striking: rather than treating an AI agent as a tool you invoke for a discrete task, a "loop" authorizes a swarm of agents to operate continuously in the background, perpetually monitoring, refactoring, and improving a project without human initiation for each action.
Why it matters
The agentic AI era has so far been defined by careful human management: set a clear goal, check in on progress, and keep the agent from straying beyond its prompt. Loops invert that model. Instead of a human kicking off each task, agents run perpetually, finding their own work and submitting it for review.
If Cherny is right that this is as significant as the shift from hand-written code to AI-assisted coding, the implications are substantial. Development teams could see a baseline of continuous, autonomous improvement layered on top of their normal workflow — architecture refinements, deduplication, and cleanup happening around the clock. The pull request remains the human checkpoint, but the labor of identifying what to fix is fully delegated.
The idea also raises immediate questions about cost, oversight, and code quality. Continuously running agents consume compute and API budget without pause. A swarm that never stops could generate a flood of pull requests that overwhelms human reviewers — or, worse, introduce subtle regressions that accumulate faster than they can be caught. Cherny's endorsement lends the concept credibility, but the practical guardrails are still being worked out.
Public reaction
No strong public signal was available from Reddit or other discussion forums at the time of this article. The concept is newly articulated in a conference setting and has not yet generated broad community discussion that could be captured here.
What to watch
- Whether other AI coding toolmakers (GitHub Copilot, Cursor, Cognition, etc.) adopt or endorse the loop model.
- How teams manage the review burden when agents submit pull requests continuously.
- Cost and resource implications of perpetually running agent swarms in production environments.
- Whether loops extend beyond code maintenance into other domains like testing, documentation, or security auditing.
- Anthropic's own productization of loop concepts within or around Claude Code.
Sources
Public reaction
No Reddit or public discussion data was available at the time of writing. The loop concept was articulated in a conference talk and has not yet generated measurable community reaction.
Open questions
- Will developers embrace perpetually running agents or see them as a review-burden risk?
- How do continuous loops affect API costs and compute budgets in real-world projects?
- Are there documented cases of loops introducing regressions faster than humans can catch them?
What to do next
Developers
Experiment with a simple loop using Claude Code or a comparable agentic tool: set one agent to continuously scan for duplicated code patterns and submit PRs for unification.
Cherny's own setup demonstrates that even a single well-scoped loop can produce useful, continuous improvements. Starting small lets you evaluate the review burden and quality before scaling.
Founders
Assess whether continuous agent loops could reduce your team's technical debt backlog, and model the API cost of running agents 24/7 before committing.
Loops could give small teams leverage by automating maintenance work, but the cost model is unproven at scale and could erode margins if not monitored.
PMs
Define clear acceptance criteria and review SLAs for pull requests generated by loop agents before deploying them in your engineering workflow.
Continuously running agents can flood your PR queue. Establishing review thresholds and quality gates prevents agent-generated changes from overwhelming human reviewers.
Investors
Track which AI coding toolmakers explicitly adopt the loop model and whether it correlates with retention or productivity gains in customer accounts.
Cherny's endorsement signals a potential platform shift. Companies that productize loops early could capture developer mindshare, but the thesis needs validation through usage data.
Operators
Audit your current CI/CD pipeline capacity to handle a potential increase in automated pull requests before introducing loop-based agents.
Loops generate continuous background work. If your review and deployment infrastructure isn't prepared for higher PR volume, the agents' output could bottleneck rather than accelerate delivery.
How to test
- 1Identify a well-scoped, repetitive maintenance task (e.g., finding duplicated utility functions or improving naming consistency).
- 2Configure a single agent to run on a loop, scanning the codebase for instances of the target pattern and submitting pull requests for each finding.
- 3Run the loop for a fixed test window (e.g., 24–48 hours) and collect all generated PRs.
- 4Review each PR manually, noting quality, relevance, and any regressions or conflicts with concurrent human work.
Caveats
- Cherny described loops in the context of Claude Code specifically; results may vary with other tools.
- Continuous agent operation can incur significant API costs — set spending limits before testing.
- A small or static codebase may not produce enough work for a loop to demonstrate value.
- Agent-generated PRs may conflict with concurrent human commits, requiring merge resolution.