Opinionated notes from an internal Zalando unconference on agentic engineering. Covers three themes: documentation and knowledge bases in an AI era, why ROI matters once tokens carry real cost, and agentic practices across the SDLC with code review at the centre. Closes on whether these are AI problems at all, or old software engineering problems resurfacing.

Some opinionated etiquettes for AI-assisted code generation, inspired by the Boy Scout Rule. Covers writing human-readable code in hybrid codebases, reviewing AI-generated code and PR descriptions, acknowledging AI in commit messages, maintaining project steering rules, avoiding test manipulation, structuring code review feedback across linters, AI agents, and humans, and evaluating AI-introduced dependencies. Based on hands-on experimentation with agentic engineering workflows, not research data.

Covers the limitations of vibe coding with Claude Code for production-grade work — context window constraints, ad-hoc design, and prompt-as-context friction — and how spec driven development with Kiro addresses them. Walks through Kiro’s three-document workflow (requirements, design, tasks) using a real Cloudflare R2 migration as a worked example. Discusses repository structure, where Kiro excels, and common pitfalls to avoid.