The way engineers interact with large language models is changing. Instead of crafting the perfect prompt and hoping for the best, a new methodology emerging from Anthropic’s ecosystem has developers building systems that prompt AI models repeatedly, letting the software observe its own output, adjust, and try again.

From prompts to loops

The shift started gaining traction after Anthropic published findings on December 19, 2024, explaining how agents could execute complex tasks using iterative cycles and environmental feedback. Rather than following static code paths, these agent loops allow models to dynamically control their own processes, observing results and course-correcting in real time.

Boris Cherny, a key figure in the development of Claude Code, captured the philosophical shift with characteristic bluntness.

“I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.”

Anthropic recommends what it calls an “evaluator-optimizer” strategy for agent workflows where clear success metrics exist. You build one layer that generates output, and another layer that judges whether that output is good enough. If it isn’t, the loop runs again.

Better results, bigger bills

Agent loop systems can result in token costs approximately 4x higher than standard chat interactions. Multi-agent setups, where multiple AI models coordinate with each other, can push that figure to roughly 15x the cost of a conventional exchange.

For context, if a company is spending $10,000 a month on AI API calls with traditional prompting, switching to multi-agent loops could balloon that to $150,000.

Loop engineering as a discipline

The term “loop engineering” was popularized around June 7-9, 2026, when endorsements from Cherny and other AI developers began framing loop design as an essential competency for working with agentic AI.

The distinction from traditional prompt engineering is meaningful. Prompt engineering is about crafting the right input. Loop engineering is about designing systems and stopping criteria, knowing when a model has iterated enough to produce a satisfactory result, and when it’s just burning tokens going in circles.

Anthropic’s original December 2024 post laid the conceptual groundwork, but the practical adoption took roughly 18 months to mature.

What this means for investors

The cost dynamics are worth watching closely. Companies that figure out how to get loop-quality results without loop-level token consumption will have a genuine moat.

It’s also worth noting what this trend doesn’t affect. Despite the rapid evolution in AI agent methodologies, there’s no direct connection between loop engineering and crypto assets or blockchain technology.