Glossary
What is loop engineering?
Loop engineering is designing the act, observe, decide, repeat cycle an AI agent runs until a stop condition, not just writing the prompt inside it. It's the full stack: prompt (what to do), context (what the agent can see), loop (the repeat logic), harness (the tools and guardrails around it). Skip the loop layer and an agent repeats its own mistakes instead of correcting them.
Why it matters
A loop is a multiplier: it turns one prompt into ten, fifty, a hundred iterations. That multiplies output, but it also multiplies whatever the agent gets wrong, and "looks done" isn't a stop condition an agent can judge on its own. The loop needs a checkpoint where a human actually looks at what got built, otherwise the review work that used to happen once now has to happen at every iteration, and it piles up on whoever's supposed to sign off.
How Patchrooms uses it
Patchrooms is that checkpoint. Embed one script tag on the preview the loop just produced, and a reviewer (no account needed) clicks the broken element and leaves a comment, typed or a voice note that's auto-transcribed. Patchrooms captures the DOM selector, a screenshot, URL, viewport, and console errors alongside it. "Copy agent feedback" turns that into a Markdown block (selector, goal, constraints, screenshot ref) the next loop iteration can read as its next prompt. An agent can also skip the copy-paste and pull it directly: the MCP server exposes list_reports and get_report so a coding agent reads the human's feedback and calls set_status itself once it's fixed.
Try the live demo