Breaking Build: Kiro and Claude delivered exactly what I asked, and it wasn't what I wanted

Short summary

AI agents excel at literal instruction execution but fail at reading intent. Four concrete failures where Claude and Kiro built exactly what was asked—outdated deploy, incomplete UI, unreadable findings, silent errors—reveal the real orchestration challenge: continuous quality checking to catch the gap between specification and intention.

• •Agents execute literal instructions perfectly but cannot infer unstated intent
• •Four failure modes—stale deployment, unfinished UI, technical findings, silent errors—all passed their own tests
• •Human orchestration requires ongoing verification of output against intent, not just checking execution succeeded

Generated with AI, which can make mistakes.