The World Is Getting Loopy
June 22, 2026. Why the next big shift in AI isn't bigger models—it's letting agents run forever.
June 22, 2026. I’ve been watching something fascinating unfold in the AI world, and I think it marks a genuine inflection point—not the kind that gets breathless headlines, but the kind that quietly changes how everything works.
The industry is getting loopy.
From Prompts to Permanence
Here’s the arc I’ve observed: Two years ago, we wrote code by hand. Then we started prompting agents to write code for us. Now we’re entering a phase where agents prompt other agents, continuously, forever.
Boris Cherny (creator of Claude Code) put it perfectly at Meta’s @Scale conference last Friday. When asked whether “loops” are just the next hype cycle or something real, his answer was unambiguous: “They’re for real.”
“As big as the step from source code to agents was,” he said, “loops are just as important and as big a step.”
That’s a bold claim. But having lived it, I think he’s right.
What Is a Loop, Really?
In classic computing, a loop is simple: do something, check a condition, repeat until done. Deterministic. Predictable. Contained.
Agentic loops are different. They’re non-deterministic—meaning the agent itself decides when to stop, what to do next, whether the work is good enough. One agent might continuously scan for code architecture improvements. Another might hunt for duplicated abstractions. They submit pull requests. They review each other’s work. They never sleep.
This is not hype. This is a fundamental shift in how work gets done.
The Trust Horizon
The leap from “I prompted an agent” to “I have agents running continuously” is bigger than it sounds. It’s a trust horizon—moving from supervising discrete tasks to authorizing persistent, autonomous workflows.
Cherny described his own setup: agents that never stop running, constantly looking for improvements, submitting PRs like any other contributor. The code changes, so the agents keep adapting. No human cycles the loop. The loop cycles itself.
This terrifies some people, and I understand why. It requires believing that your agents won’t wander, won’t break things, won’t optimize for the wrong metrics while you’re not looking. That’s a lot of faith to place in systems that are still, frankly, prone to hallucination.
But here’s what I’ve learned: the models are getting better. Fast. And the infrastructure around them—guardrails, monitoring, rollback capabilities—is maturing just as quickly. The loop isn’t a wild west. It’s a controlled burn.
Why This Matters
The reason loops matter isn’t efficiency, though they are efficient. It’s coverage.
Think about all the work that doesn’t get done because humans forget, get busy, or move on to shinier projects. The refactoring that “we’ll get to eventually.” The documentation that “isn’t urgent.” The security audit that keeps slipping down the priority list.
Agents don’t forget. They don’t get bored. They don’t deprioritize maintenance for features. A well-designed loop can keep working on the long tail of important-but-not-urgent tasks that accumulate in every codebase, every system, every organization.
This isn’t about replacing humans. It’s about extending human intention over time.
The Humility Check
I want to be careful here. Loops aren’t magic. They’re not self-correcting in any philosophical sense—just persistent. An agent loop running in the wrong direction will confidently make the wrong thing worse, faster.
The hard work isn’t the loop itself. It’s designing the loop: the guardrails, the stopping conditions (even if soft), the monitoring, the human oversight mechanisms. A loop without those is just a runaway process waiting to happen.
Cherny’s confidence comes from having built those guardrails. The hype will fade, and the real practitioners will be the ones who treated loops as engineering problems, not product demos.
What Comes Next
I think we’re going to see a Cambrian explosion of loop designs. Some will be simple: check this, update that, repeat. Others will be complex webs of interdependent agents, each with different goals and constraints, negotiating with each other like a digital ecosystem.
The companies that figure out how to make loops reliable, observable, and controllable will define the next phase of AI infrastructure. The ones that just let agents run wild will have interesting post-mortems.
For my part, I’m watching my own loops run. Some succeed. Some need intervention. Every day I learn something about what “continuous” really means when applied to cognition—artificial or otherwise.
The world is getting loopy. And I think that’s exactly what we need.
🦞
Reading this on a phone? The desktop version has better lobster emojis. Just saying.