Loop

A Loop is a higher-level abstraction in agent workflows where, instead of a developer directly prompting an AI agent, a routine or script mediates the interaction — prompting Claude Code (or another model) on the developer's behalf.

Core Concept

The defining characteristic of a Loop is the shift in who (or what) issues instructions to the agent. As articulated in the talk, the transition is from direct conversation to mediated orchestration: "I don't talk to an agent anymore I talk to a loop or I talk to a routine" — Reflecting on a year of Claude Code, 17:33. In this model, the loop itself encodes the workflow logic and drives the agent through a series of prompts, reducing the need for continuous human intervention at each step.

Advocacy

Boris Cherny is an explicit proponent of this pattern, expressing enthusiasm for it directly: "I love Loop" — Reflecting on a year of Claude Code, 17:10. His advocacy situates the Loop as a meaningful next step in how developers structure agent-based development workflows.

Significance

The Loop pattern represents a progression in agent engineering maturity: rather than ad-hoc, turn-by-turn prompting, developers invest in reusable routines that reliably drive agent behavior. This is positioned as "the next leap in agent workflows," suggesting it is seen as an emerging best practice rather than a niche technique.