The AI World is Getting 'Loopy'

What Are AI Loops?

At Meta's @Scale conference, Claude Code creator Boris Cherny introduced the concept of "loops" as the next evolution in AI development. Unlike traditional agentic AI that requires human oversight for discrete tasks, loops authorize swarms of AI agents to work continuously in the background—endlessly.

Key Technical Insights

The Evolution of AI Development

Cherny outlined the progression:

Two years ago: Developers wrote source code by hand
Current state: AI agents write the code
Next phase: Agents prompt other agents that then write code

"As big as the step from source code to agents was, loops are just as important and as big a step," Cherny explained.

How Loops Work in Practice

Cherny shared specific examples of loops running continuously in his own work:

Architecture improvement agent: Constantly looks for ways to improve code architecture
Abstraction unification agent: Searches for duplicated abstractions that can be unified
Both agents submit pull requests autonomously like human developers
Because code is constantly changing, these agents never stop running

Technical Foundation

Connection to Classical Computing

Loops aren't entirely new—they build on recursive loops from intro computer science:

Traditional recursive loops call themselves to repeat actions with a stopping condition
AI loops use non-deterministic logic: a subagent decides when to stop instead of a predefined condition

The Ralph Loop Pattern

One popular implementation is the "Ralph Loop" (named after Ralph Wiggum):

Summarizes all work the model has completed
Asks if the goal has been accomplished
Bounces the model back and forth until the task is complete
Prevents AI models from getting lost during long-running tasks

Test-Time Compute Strategy

Loops align with OpenAI researcher Noam Brown's observation:

Contemporary models can solve nearly any problem with enough compute
Loops keep throwing compute at problems until they're finished
Particularly effective for hill-climbing problems like improving codebases
Models make incremental improvements until reaching a threshold—or indefinitely

Business Considerations

Cost Implications

Token consumption is significant:

Like agentic AI, loops burn through tokens faster than Q&A chatbots
Since loops run continuously, there's no ceiling on spending
Fine for token-selling companies like Anthropic
Potentially expensive for end users

Value Proposition

Despite costs, benefits could be substantial:

With proper oversight of token spend, drift, and AI issues
For the right problems and setups
Returns could outweigh operational expenses

Industry Significance

This represents a fundamental shift in how developers interact with AI:

From writing code manually
To managing AI agents writing code
To AI agents managing other AI agents that write code

The loop paradigm places significant trust in AI systems but could be the breakthrough needed for AI to handle real, continuous production work.