How I actually ship with AI agents

I use agent tooling daily across OpenClaw, Claude Code, and Lovable. The productive pattern is narrow scope, explicit contracts, and verification I control.

The loop

• Write the acceptance criteria before the prompt — if I cannot test it, I cannot delegate it
• Give the agent one file surface at a time, not the whole repo
• Review diffs, not summaries
• Keep infrastructure and secrets decisions human

Treated as a fast pair with no memory of consequences, agents are a real multiplier. Treated as an authority, they quietly generate technical debt at machine speed.