Humans and Agents in Software Engineering Loops

Kief Morris argues that the real promise of AI agents in software development is not full autonomy or human micromanagement, but building “working loops” that reliably turn ideas into outcomes. Humans should design and steward these loops: clarify goals and constraints, decompose work, define quality signals and guardrails, integrate tools, and monitor performance. Agents sit inside the execution subloops (e.g., drafting code, tests, docs), while humans own intent, prioritization, evaluation, and continuous improvement. Success comes from strong feedback signals and observability, versioned prompts and artifacts, safety checks, and iterative refinement of the loop rather than the outputs alone.

Question: In your team’s delivery loop, where would an agent add the most leverage—code generation, test creation, incident response, or planning—and what measurable signals would you use to confirm it’s improving outcomes?