Agents write code. They don’t do software engineering.
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Long-running coding agents can now submit pull requests autonomously, but the real question isn't how much to supervise them — it's which categories of work they should own. Writing code is pattern recognition that LLMs excel at, while software engineering involves trade-offs, business context, and judgment that agents lack. A three-tier framework is proposed: Tier 1 (agent-led, developer-reviewed) for boilerplate and scaffolding; Tier 2 (agent-assisted, developer-guided) for feature work within known domains; and Tier 3 (developer-led, agent-supported) for architectural decisions and cross-boundary changes. Over-allocating to agents causes costly rework; under-allocating wastes developer time on mechanical tasks. Teams that audit their work distribution — not just their agents — reclaim developer time for high-judgment work that agents cannot yet handle.
Table of contents
Code writing and software engineering are not the same jobWhat developers actually ownWhere agents deliver the most valueA three-tier model for dividing the workThe cost of getting the split wrongAudit the work, not just the agentsSort: