How is it different from “human-in-the-loop”?
Classic human-in-the-loop puts a person at the decision point: the machine proposes, the human approves each step. An AI-in-the-loop workflow inverts the volume. The agent carries a goal across a dozen edits, runs the tests, and presents a finished end state; the human’s leverage moves up a level to designing the loop itself — the conventions, the constraints, and the graders that decide whether the work is acceptable. You stop reviewing every line because at the volume agents produce, line-by-line review is the bottleneck that stops the practice from scaling.
What keeps it trustworthy?
Conversation alone has a hard ceiling: prose context competes with ten thousand training examples pulling the other way and degrades over a long session. So the durable levers are structural — the compiler and the environment constrain the agent, and read-only grader agents (which physically cannot edit the code they judge) check the output. The human governs that structure rather than babysitting each step. That is what “in the loop” means here: present in the design and the validation, absent from the keystrokes.
Where does the human add the most value?
In direction and judgement: choosing what to build, encoding the unusual rules a model will never infer from public code, and deciding what “good” means so a grader can enforce it. The role shifts from writing code to governing the systems that produce and grade it.
The full method — why conversation plateaus, how to constrain agents with the environment, and how the human role changes — is in the pillar guide: agentic software development, as I actually practice it.