"Outside the Loop" is a collaboration mode in which humans specify only the outcome requirements and delegate all implementation details to AI agents; it is also known as vibe coding.
The mode with the least human involvement in Bockeler's three-mode classification. You tell the agent "I want this feature," then leave everything to the agent and simply receive the finished product. This is essentially the same concept as vibe coding, a term coined by Andrej Karpathy. ### What's the Problem? Something that works gets built quickly. However, "working" and "correct" are two different things. In Outside the Loop, humans don't review the architecture or dependencies chosen by the agent, allowing inefficient implementations to quietly accumulate. Bockeler describes this as a "spiral of inefficient solutions." The typical pattern goes like this: the agent adds an unnecessary library to solve a problem. The next instruction introduces a compatibility issue with that library, and yet another workaround gets added. Each individual change seems reasonable in isolation, but technical debt as a whole grows at an accelerating pace. ### When Outside the Loop Is Acceptable Outside the Loop is not a problem in situations where long-term maintenance is unnecessary, such as creating prototypes or throwaway scripts, or experimental coding for learning purposes. Conversely, it is a mode to be avoided for production code that a team continuously maintains.


"On the Loop" is a collaboration mode that focuses on improving the harness (operating environment, constraints, and tools) rather than individual outputs of AI agents, and represents the recommended human position in the practice of harness engineering.

"In the Loop" is a collaborative mode in which humans review and correct AI agent outputs one by one. While quality control is reliable, it tends to create a bottleneck where human review cannot keep pace with the agent's generation speed.

Vibe Coding is a general term for a development style in which developers communicate their intent in natural language and AI handles code generation and modification; it was selected as the "Word of the Year" for 2025.
