Type: url
Original: https://developers.cyberagent.co.jp/blog/archives/62404/
Status: generated
Summary
A CyberAgent Developers Blog post by a Unity engineer at GOODROID comparing two attempts to move the development of internal Unity client infrastructure to spec-driven development with AI agents, so that humans wrote no code. In the first, a single giant spec that took about 73% of the context window led to frequent compaction, skipped workflow steps, and a loop in which fixing the flow with more rules made it worse. The second split the work into 21 specs, minimized always-loaded context, orchestrated Claude Code with Codex across providers, enforced tests through deterministic gates in hooks and GitHub Actions, and ran agents without permission prompts in an isolated Docker environment, after which the volume of pull requests became the new bottleneck.