The Midnight Garden · Learning guide
This is an honest map of how AI assistants actually work: eight interface styles, five levels of build complexity, the skills behind each one, and what genuinely changes the cost. One difference from every other guide on this subject -- the proof is on the page. The chat bubble in the corner is a working agent, and you can test every claim below against it, right now.
The exhibit
t.ROY -- lower right -- is an L5 on the scale above: it knows which page you are reading, it can check a real calendar and offer real meeting times, it saves nothing about you without a consent card, and every action it takes leaves the owner a receipt. It is the same system The Midnight Garden deploys for clients, wearing its own brand. Try the claims before you read the theory:
Conversations auto-delete after 30 days. Nothing is saved about you without a consent card you can read. That policy is part of the lesson.
Interactive guide
Complexity map
Learning notes
Grounding examples
If a pattern here looks close to something in your own organization, map the workflow before choosing a tool. That is where we start every build.
Full disclosure
Ask the vendor to show the system unsure, over budget, and offline. A serious build has a degraded mode that still helps ("I'm limited right now -- leave your info"), never a blank widget. Ask this page's agent what its failure mode is; it will tell you.
"Fully autonomous" is a red flag, not a feature. The trustworthy pattern: the assistant acts, the owner approves anything that touches the calendar or the outside world, and every action is logged. You stay in command of your own business.
Every assistant gets things wrong. The question is what happens next: mistakes should be logged automatically, reviewed on a schedule, and FIXED only with the owner's sign-off -- improvement with receipts, not an AI quietly rewriting itself.
The chat bubble is half the product. The owner needs leads, booked meetings, unresolved questions, costs, and a daily brief that stays quiet when there is nothing to say. If the demo never shows the owner's side, there isn't one.
How long do conversations live? What gets saved, and with what consent? A vendor who cannot answer in one sentence has not thought about it. (This page's agent: 30 days, consent cards before anything is stored, no exceptions.)
Calendar, CRM, payments, phone, email -- each integration moves the cost and the risk. Make the list visible before anyone quotes a number, and be suspicious of anyone who quotes one without asking what the assistant must touch.
The Midnight Garden
If a workflow in your business looks like one of these patterns, the conversation about what it would actually take is free -- and you can start it with the agent itself, right now. It will check the real calendar.
Or just ask it, lower right. It knows you made it to the bottom of the page.
