Tips for writing software with LLM agents (liveandletlearn.net)

A senior software engineer shares ten practical tips for “agentic engineering” — using LLM agents (e.g., Aider, Zed, Cursor with your own LLM key like Anthropic) as integrated, in-editor pair programmers rather than separate chat windows. The post argues that agents accelerate creative, higher-level work (system design, UX, trade-offs) and lower the friction for refactors and small improvements, while also making immediate learning easier by explaining patterns, APIs, and alternatives. Key tooling and workflows highlighted include editor-native agents (Zed), commit-centric tradeoffs (Aider), and the memory-bank/.ai-instructions pattern to persist project context and avoid repetitive prompts or sycophantic replies. Critically, the author warns of real pitfalls and gives concrete practices: always own and personally review every generated change before committing; iterate for readability, error handling, and focused tests; start sessions with full context to mitigate token limits; run pre-implementation design discussions with the agent; and plan small, reviewer-friendly diffs instead of sweeping edits. The post positions agents as powerful collaborators that boost productivity and learning but stresses developer responsibility — you remain the decision-maker, reviewer, and maintainer, so don’t commit code you can’t defend.