The Dashboard Was Always a Prompt

Ron Reynolds · 2026-09-22 · 6 min read

You know the sentence before the first click.

"Send the spring win-back to everyone who hasn't bought since March, and show me the list before it goes." That's the whole instruction. It's in your head, complete, at 7:40 in the morning with the coffee still hot. Then you open the marketing suite. Campaigns. New. Type: email. Audience: build a segment. Condition: last order date, before, pick a day from a calendar widget that opens on the wrong month. Save the segment. Name the segment. Back to the campaign. Attach the segment. Subject line. Template. Schedule. Review. Somewhere around click nine the sentence in your head has been fully dismantled and reassembled in the vendor's order, and you press the button.

Now watch what happens when you pay someone to automate that screen for you. They write a script, and the script reads like this: go to Campaigns, click New, choose email, open the segment builder, set the field to last-order-date, set the operator to before, set the value to March 1, save, name it, go back, attach, fill the subject, choose the template, schedule, confirm.

Read that back. It's a prompt. A paragraph of imperative sentences addressed to a system. The vendor wrote it years ago and shipped it as a screen, nouns and verbs pinned in place, blanks left where your particulars go. Every time you used the dashboard you were reciting the vendor's sentence and filling in the blanks by hand. You were the thing that moved the mouse so the sentence could get spoken. We were meat puppets, and the strings ran up to a grammar somebody else wrote.

Here's the part that took me longer to see. Agents do no better with dashboards than we do. Hand an agent that same screen and you get a browser-driving contraption that takes a screenshot, finds the button, clicks at the coordinates, waits, takes another screenshot. A puppet driving a puppet. Give it the API instead and you've removed the pixels; the sentence underneath is the same one, and the agent still has to be taught the vendor's grammar from a spec before it can say a word. The direction of translation never moved. Caller learns callee.

So the dashboard is a foreign language to both of you. The owner thinks in the sentence from 7:40. The agent, as it happens, thinks in sentences too. That is the entire reason it's useful.

ComOS takes the dashboard out and lets the two of you talk to the system directly. That can sound like a chat box bolted onto the old software. The mechanics run the other way. The system speaks first. Every capability it has is announced in prose, the way you'd describe a job to a new hire: "Enter a tenant to receive its tool surface. The gateway is a small catalog. List tenants, then enter one. The reply carries the entered platform's real action tools with descriptions and schemas." Your agent reads that the way you would read it. Then it says what you want, in the words you gave it, and the system does it.

And the system talks back, in sentences, including when it says no. Ask for something your balance won't cover and the answer is a paragraph: "Your Com balance is below this act's price. Nothing was charged and the act was not run. Top up via your funding rail, then retry. Read prices free anytime." Name a tenant you don't own and it tells you to enter one you do. When it says no, it says what happened and what would make it yes. A dashboard's version of that is a red banner and a support ticket.

There is no login page. The addresses where a dashboard would live all point at one page whose center is a single paragraph, and the paragraph is an instruction you paste into whatever agent you already use: read the disclosure, connect, admit me, show me what I can build, create my first tenant. Across the entire public site there are three forms. One takes a URL and scans your website for agent-readiness. One takes a card, for a customer paying for something. One searches the blog. Nobody runs a business on ComOS by filling out a form, because there isn't one to fill out.

Dashboards do one thing well, and that thing survives: showing you state at a glance. On ComOS you ask for it. "This week's orders against last week's, by channel." A table arrives, built for that question, and it's gone by lunch. The picture is still useful. You just don't live inside it anymore.

The login page was also the dashboard's whole idea of authority: you're in or you're out. That idea moved to the door and got more precise. You authorize your agent once, with OAuth, and every sentence it speaks to the system carries that authorization with it, scoped to what you allowed. Meanwhile the routine traffic of commerce (prices, stock, availability, who paid whom) crosses the network as pure data, with no room in a number for a command to hide. Anyone may hand your agent data. Only you hand it instructions. The screen is gone; the lock moved to the door.

For a long time I thought the forms were the software. They were a translation layer, and we got so fluent in it that we stopped noticing we were translating. The sentence at 7:40 was always the real instruction. If you want to try speaking it to the system, this is the one to hand your agent:

Further reading Prompts Are the Agentic Language — the first half of this idea: the machine learned the owner's grammar, and who may speak it is the security model. The Relationship Is a Prompt — how a whole business states its network posture in a paragraph.