We timed our AI agent on four real build tasks. Here's every number.

We built Mavi to handle the part of the platform people were quietly hating: learning where every setting lives before they could get anything done. You describe what you want in a chat, and Mavi does the building itself, creating the blocks, writing the copy, setting prices, linking the steps, checking its own work.

Demos are easy. We wanted to know if it holds up on real work, so we gave it four tasks, one line each, no follow-up coaching, and timed every one against what it would have cost us by hand.

The prompts:

• "Build a booking funnel for a barbershop haircut, with a $10 prepayment to hold the slot"
• "Build a website for a coffee shop called 'Aroma' with a menu (latte, cappuccino, desserts) and a map showing the address"
• "Build an online store for handmade candles: a lavender scented candle, a candle in a glass jar, and a set of 3 candles"
• "Build an online course called 'Photography basics for beginners' with 3 modules and paid access"

The funnel took 2 minutes 7 seconds. It came back as a real multi-step flow: a welcome message, a book-a-haircut trigger, a step to collect the preferred time, a step for the name, a phone field with input validation, the $10 payment step, an "I've paid" branch that ends in "your slot is confirmed", and a separate cancel branch. Doing that by hand is 20 minutes of clicking, minimum. While it worked we could watch it call the builder functions one at a time, six blocks, two connections, a check, a fix. Not a template dump, the same steps we would have taken.

The candle store took 35 seconds. We asked for 3 products and got 6, across 4 categories, each with a description, a price between $18 and $39, and a photo. Mavi apparently decided our imaginary candle brand needed a Gift Sets category, and honestly, it was right. Filling out a product catalog by hand, writing the descriptions, picking the prices, laying out the page, is an afternoon of work.

The course took 2 minutes 12 seconds: 3 modules in a sensible order, 6 lessons with real titles like "The Exposure Triangle", homework on each one, a paid plan attached. Our own course-completeness checklist hit 100 percent, which is more than we can say for most courses built by actual humans. That is a two-hour job normally.

The website is where we'll be honest. It took 2 minutes 27 seconds, and mid-build Mavi caught its own broken nav links (the buttons pointed at sections with no anchors yet) and fixed them without being asked, which was a nice moment to watch. But the output is a scaffold, not a finished site. The copy reads like a coffee shop described by someone who has heard of coffee, the layout is plain, and you still owe it an hour in the editor. For a website, it saves you the setup, not the polish.

You talk to Mavi the way you'd talk to Grok, Claude, or GPT: describe the task in plain words and it goes and does it. The difference is it's acting inside a live project, not handing you text to paste somewhere. Nothing happens behind a curtain either. It shows a plan, waits for you to approve anything it can't undo, and logs every change. It will also make calls you didn't ask for, like the three bonus candles, so review before anything goes live. Think of it as an eager junior who ships fast and occasionally needs a second look.

The pattern we're seeing: the more repetitive the setup, the bigger the win. A product catalog or a course structure is where it pays for itself. A one-off landing page, less so.

For anyone who's tried agent builders: after the agent finishes, where does your time actually go? Editing the copy, fixing the layout, or rewiring the logic it didn't get right?