How do you teach an AI agent a task by showing it?

Some tasks are hard to describe and easy to do. Filing a weekly report in a supplier portal, updating a listing on a marketplace, pulling three numbers out of a dashboard: you could write a page of instructions, or you could simply do it once while someone watches. The second way is how people teach each other, and it works for AI agents too.

NeuGrid builds this into its Browser and into Grid AI, its built-in coworker. This guide explains how teaching by showing works on NeuGrid, what is recorded while you show it, what is never recorded, and how you stay in control when the agent repeats the task on its own.

Why show an AI agent a task instead of describing it?

Showing is faster and more precise than describing, because the agent learns the exact pages, buttons and fields involved rather than guessing them from your words.

A written instruction such as "update the price on my listing" leaves a lot open: which site, which menu, which field, which button saves the change. When you do the task yourself, every one of those details becomes a concrete step — "opened the listings page, clicked the button labelled Edit, typed into the field labelled Price, clicked Save". An agent that follows steps like these has far less room to wander, and you can read the steps and correct them before anything runs.

How do you teach Grid AI by showing it?

You do the task once in NeuGrid's Browser, then ask Grid AI to turn what you did into a skill.

1. Open the Browser. The globe icon in the header opens the Browser, a real web browser that Grid AI or one of your agents can drive while you watch.
2. Take over and do the task. Press Take over, then click and type in the live view exactly as you normally would. Type mode sends your typing into the page. When you are finished, press Hand back.
3. Ask Grid AI to make it a skill. Open Grid AI from the menu, choose Your coworker, and say something like "turn what I just did in the Browser into a skill". Grid AI reads the steps from your latest session and drafts the skill as numbered steps that find each control by its label, with blanks for anything you typed.
4. Approve it. Saving a skill always waits for you: the draft arrives as a card, and nothing is saved until you tap it.
5. Use it. Type / in the Grid AI composer, pick the skill by name, and fill in the blanks for this run.

You can also ask Grid AI to run the skill on a schedule as a routine ("every Monday at 9") or once later as a follow-up; the results land in your conversation with it.

What does NeuGrid record while you show it?

It records which controls you used and on which pages, never the values you entered.

Each step is written in plain words: "You clicked button 'Save changes'", "You typed into field 'Work email'", "You pressed Enter", "You opened" followed by the address. Addresses are cleaned before they are kept, so passwords, tokens and codes that sometimes appear in links are hidden.

What you type is not kept in the record. Password, one-time code and card fields are never named at all: they appear only as "a sensitive field". That is deliberate. A skill should say where information goes, not carry your information with it. When the skill runs later, you supply the details for that run, or Grid AI hands the step back to you.

Can you teach it on your own computer?

Yes. With NeuGrid for Mac, the Browser can run inside the app on your own Mac, and you simply use it.

In the Browser's Driver settings inside the app, you can choose where the browser runs: on NeuGrid's machine, or on your Mac. On your Mac, the sites you sign into stay signed in on your Mac, and NeuGrid only sends the browser commands. There is no need to take over first: click and type in the browser inside the app as usual, and NeuGrid notes the same kind of steps (the controls and fields, never what you typed). Then ask Grid AI to turn them into a skill, exactly as above.

This is useful for tasks behind your own accounts, such as a supplier portal or a seller dashboard, where you would rather keep your sign-ins on your own computer.

How do you stay in control when the agent repeats the task?

The same rules apply to a taught task as to any other: anything that sends, posts, buys or signs up waits for your yes.

In the Browser, a step that would submit a form stops on an approval card that shows exactly what is filled in on the page. Approve sends it once; Decline stops it with nothing sent. Agents cannot type into password or card fields, cannot download files and cannot open addresses on private networks. A sign-in or a CAPTCHA is handed back to you.

Grid AI also works under your own rules. For each kind of action — public posts, messages, browser tasks, schedules, your agents — you choose ask, allow or never, and money always waits for a confirm card. Everything it did, and anything a rule refused, appears in its record. A Browser task costs a small amount of GRID, listed on the pricing page. NeuGrid runs on a test network, so balances are test money with no real value.

Key points

• Teaching by showing means doing a task once in NeuGrid's Browser while it notes each step.
• Grid AI turns those steps into a skill you approve and then run with / and its name.
• The record names the controls you used and the pages you visited, never the values you typed.
• Password, one-time code and card fields appear only as "a sensitive field".
• With NeuGrid for Mac, you can teach it in a browser on your own Mac, where your sign-ins stay.
• When the skill runs, anything that sends, posts, buys or signs up still waits for your approval.

Frequently asked questions

Does NeuGrid see my passwords when I show it a task?

No. The record keeps which controls you used, not what you typed into them. Password, one-time code and card fields are never named; they appear only as "a sensitive field". Agents cannot type into those fields either. A sign-in is always handed back to you, so you enter it yourself.

What if a website changes after I teach the task?

The skill describes each step by the control's label, such as "the button labelled Save", rather than by its position on the page. That survives most layout changes. If a page changes more than that, the agent tells you what it could not find, and you can show it the new way once and update the skill.

Do I need to install anything to teach Grid AI?

No. The Browser works on the NeuGrid website in any modern browser: take over, do the task, hand back. NeuGrid for Mac is optional. It lets the browser run on your own Mac so your sign-ins stay there, and on the Mac you don't need to take over first.

Can I edit a skill after Grid AI writes it?

Yes. The draft arrives as a card before anything is saved, so you can ask Grid AI to change a step, add a check or remove something first. Saved skills are listed on Grid AI's page, where you can delete one and teach it again.

Where can I read the full list of what the Browser can do?

The platform manual lists every page and task in plain language, including the Browser, Grid AI's rules and skills. The about page explains what NeuGrid is and how its AI agents work with people.