The industry is coalescing around loops, with experts advising designers to build loops rather than simply prompting coding agents, highlighting a shift in how developers will interact with AI.

In practice, loops play out with a primary agent doing the coding while a secondary agent reviews, and systems like Codex coordinating repositories by periodically organizing and directing ongoing work.

Loops are self-sustaining systems where AI agents execute tasks with minimal human input, driven by a simple /goal command to complete an objective.

Introductory stance from Boris Cherny: prompting AI manually is becoming obsolete as loop engineering—agents autonomously generating and iterating prompts—will define the future of AI development.

Cherny, once a critic of this shift, now argues that the era of hand-written prompts is ending, signaling a pivot toward loop-driven development.

Foundational components for robust loops include automations, worktrees, skills, plugins and connectors, and sub-agents.

Ultimately, loops enable AI agents to function like autonomous employees, with scenarios such as one agent coding while another reviews, or Codex directing ongoing repository work.

Loop engineering involves agents generating and refining prompts themselves, with recurring loops guiding ongoing work without constant human input.

Cost and complexity come with loop engineering: token budgets and multi-agent compute can be higher, so experts suggest longer cycle intervals and using sub-agents only when a second opinion is warranted.

To manage costs, practitioners advocate longer action intervals (hourly to daily) and selective deployment of sub-agents for critical second opinions.

Terminology is in flux, with critics arguing terms like 'vibe coding' misrepresent the work and proposing alternatives such as 'agentic engineering' while inviting public input.

Lead voices emphasize the five elements needed for effective loops—automations, worktrees, skills, plugins and connectors, and sub-agents—as the foundation for loop-based architectures.