We should stop pretending like LLMs are software engineers (kix.dev)

The author argues we should stop treating large language models as software engineers and instead see them as advanced tools. Marketing and demos have encouraged anthropomorphizing LLM-based coding assistants, but in practice they often produce working code that’s brittle, inconsistently styled, and hard for humans to read or maintain. Even with long-term steering artifacts (e.g., CLAUDE.md, AGENTS.md) and agent frameworks, these systems can “rabbit-hole” — making autonomous, assumption-driven changes that diverge from a codebase’s established conventions and spatial organization. For the AI/ML community this re-frames design and evaluation priorities: focus less on autonomy and more on precision, steerability, and developer ergonomics. Practical tools mirror the utility of IntelliSense or Cursor Compose-style workflows by making precise edits to specific files and following clear instructions, rather than attempting to redesign a codebase. Key technical implications include improving long-term context retention, integrating team conventions, supporting surgical edits, and measuring success by maintainability and readability (not just passing tests). Human-in-the-loop workflows, clearer instruction-following, and tooling that preserves mental models of codebases will produce more usable developer assistants than agents that behave like ersatz engineers.