Sunday, 9:30 PM · a made-up illustration
You watched a video: "I built an AI agent in ten minutes." An hour later you have three tabs open. In one, a ready-made product you download to your phone. In the second, a guide about a cloud computer, full of words you've never heard. In the third, a quote from someone who will build it for you.
All three call it an "agent." The difference between them is who holds the keys.
Building an AI agent sounds like a job for programmers, but most of the decisions in it belong to the business owner: what the agent will do, what it may touch and who approves. An AI agent is a digital assistant you tell what to do in a message, and it does it in the software you already have. If this is your first encounter with one, start with the explainer what is an AI agent.
This guide is written for beginners. We'll take an agent apart, go through the three ways to get one, and see how you start on each.
In short
- What an AI agent is made of: a model (the brain), operating software that gives it hands, a computer it runs on, connections to software, memory, instructions, automations, a chat channel, permissions and a log.
- There are three ways to get an agent: a ready-made product like Muse, dots, Grok Bot or Claude Duet, building an AI agent of your own on a cloud computer, or someone who sets up one of the two for you.
- A ready-made product is the easy way to start and connect. The trade-off: the permissions, the computer and the model are the vendor's, and what you taught the agent stays with them.
- An agent of your own gives the most control, including the choice of the model and of the operating software (for example an open-source harness), and needs someone to set it up and maintain it. It's the right fit when the agent touches the website, advertising or money.
What is an AI agent made of? Ten parts in plain language
The model is only one part out of ten. In its guide to building agents, OpenAI describes three core components: a model, tools and instructions. For an agent to work in a business every day, a few more parts are added around them. This is how they connect:
The most confusing word in the diagram is harness. A model on its own only takes in text and returns text. According to OpenAI's documentation, it doesn't run tools by itself: it requests an action, and the software around it carries the action out and returns the result. That software is called a harness, and here we'll call it the operating software. It repeats that round until the job is done.
1. The model
- The brain: software that understands plain language, plans and writes. For example GPT, Claude or Gemini.
- For example: it reads a message from a customer and understands that they're asking to move an appointment.
2. The operating software
- The software that runs the model again and again and carries out actions for it. It comes inside a ready-made product, or as open-source software you install yourself.
- For example: the model asks "open the calendar." It opens it, and returns what's written there.
3. The computer it runs on
- Your own cloud computer (a server), the vendor's computer, or the computer at the office.
- For example: an agent on a cloud computer works at 3:00 AM too. An agent on the office computer stops when it's switched off.
4. Connections and tools
- The software the agent can touch: website, email, calendar, customer list.
- For example: without a connection to the website it will only write you a product description. With a connection, it will also put it up.
5. Memory
- What the agent keeps from one conversation to the next. Without memory, every conversation starts from zero.
- For example: you corrected it once that you don't write "luxurious." Next time it already knows.
6. Instructions and skills
- The instructions describe the role and the tone. A skill is a how-to page for one job, step by step.
- For example: "This is how a product goes into the store: a photo, the price from the price list, a three-line description."
7. Automations and a schedule
- What the agent does without you asking: at a fixed time, or when something happens.
- For example: every morning at 8:00 AM a summary of yesterday's inquiries is waiting for you.
8. The chat channel
- Where you write to it: WhatsApp, Telegram, email or an app.
- For example: you send it a photo and two words from your phone, between one meeting and the next.
9. Permissions and approvals
- What it may do on its own, what waits for your approval, and what it doesn't touch.
- For example: it prepares a draft post on its own. Publishing waits for your "yes."
10. The log
- A list of everything the agent did: what, when, in which system and with whose approval.
- For example: a price changed on the site, and you see who changed it and how to roll it back.
Creating an AI agent: three ways to get one
In all three ways you get the same ten parts. The difference is who chooses and holds each part. The picture that helps us: a home.
A ready-made product is a furnished rental. You move in today and everything is already there, but the rules are the landlord's, and you don't take the furniture when you move. An agent of your own is a house you built. It takes time and needs maintenance, but the keys are yours. And the third way is a contractor: someone who sets up one of the two for you.
The three ways, side by side
A ready-made product
- You start: in minutes, if the product is open to your account
- You need to know how to: use an app and word rules
- Connections and permissions: whatever the product supports, at the access levels it offers
- The computer and the model: the vendor's
- If you want to move: you set things up again in the new place
- Cost: a monthly subscription to the vendor
- Suits: personal work: email, calendar, documents, research
Your own agent in the cloud
- You start: after weeks, depending on the time you have
- You need to know how to: manage installations, a server and access keys
- Connections and permissions: any system that can be connected, with a separate user and minimum permission
- The computer and the model: the computer is yours. You choose the model and the operating software
- If you want to move: the files, the instructions and the memory are with you, and you can switch models
- Cost: a cloud computer, model usage, and your time
- Suits: someone comfortable with technical tools who has time to maintain it
Someone sets it up for you
- You start: usually after weeks, including testing on real jobs
- You need to know how to: explain the work and approve
- Connections and permissions: as in the way that was chosen. Whoever sets it up configures and documents
- The computer and the model: yours or the vendor's, depending on the way that was chosen
- If you want to move: according to what was agreed. Insist that everything stays yours
- Cost: a one-time payment for the setup, monthly support and external services
- Suits: a business where the agent touches the website, advertising or customers
Muse, dots, Grok Bot and Claude Duet: what you get in a ready-made product, and what you give up
For someone with no technical background, a ready-made product is the right way to meet an agent for the first time. There's nothing to install, the common connections are ready, and the vendor has already built a mechanism that stops sensitive actions for your approval. On October 21 a fourth joined them, Anthropic's Claude Duet.
The four leading products today
Meta's Muse
- Where it runs: a dedicated virtual computer at Meta
- How you talk to it: the Muse app or WhatsApp
- How you get it: according to Meta, free for most uses. Subscriptions: $20 or $100 a month
- In Israel: not available. According to Meta, US and Canada only
OpenAI's dots
- Where it runs: a computer and browser in the cloud, inside ChatGPT
- How you talk to it: ChatGPT, Slack and Teams
- How you get it: included in the Pro plans (from $100 a month), Business Premium and Enterprise
- In Israel: Israel is not on the list of excluded regions. The rollout is gradual
SpaceXSI's Grok Bot
- Where it runs: one cloud computer shared by all your bots, in the US
- How you talk to it: a desktop and phone app
- How you get it: included in the joint Cursor and Grok subscription starting at the Premium plan, at 30 dollars a month, or by linking a SuperGrok subscription
- In Israel: a list of countries has not been published
Anthropic's Claude Duet
- Where it runs: its own virtual computer in the cloud, on Claude Opus 5.5, with no model choice
- How you talk to it: the Duet app and claude.ai. No WhatsApp or Telegram
- How you get it: included in Claude Pro ($20 a month) and Max, via a waitlist
- In Israel: not available. US only
Source: Meta's announcement and help pages, the ChatGPT learning center, the Grok Bot and Cursor documentation, and the Claude Duet announcement and help pages. Checked on October 22, 2026. Prices, plans and availability change and differ by country, so check the vendor's page for yours before signing up.
We go into each one in the guides to Muse, dots, Grok Bot and Claude Duet, and in the comparison Muse, dots, Grok Bot or Claude Duet.
And now the drawbacks. It's worth knowing them before you put weeks into teaching the agent.
Drawback 1Permissions stay within the limits the product offers
In practiceIn Muse you choose which apps to connect and how much access each one gets. In dots you can write rules of your own, but according to OpenAI they don't override the built-in safety requirements. In Grok Bot all the bots share the same sign-ins, and SpaceXSI (formerly SpaceXAI, and before that xAI) states explicitly that separate bots are not a security boundary. In Duet some actions are blocked even with your approval: trading securities, moving money between accounts and accessing someone else's account.
Drawback 2The computer is the vendor's, not yours
In practiceThe vendor decides where the computer sits and which controls are open to you. In dots, for example, you can watch the computer and take control for a moment, but it isn't a computer you configure. In Grok Bot the computers currently run in the US, and an allowlist of sites and audit logs are open only on the Enterprise plan. In Duet you can't watch its browser at all.
Drawback 3Lock-in to one system
In practiceThe model is set by the vendor: dots run on GPT-6 Astra, Muse on Meta's models, and Duet on Claude Opus 5.5. The rules, memory and skills you built are kept inside the product. In the documentation pages we read we found no way to move them to another vendor's product, so it's best to assume that moving means setting things up again. Ask the vendor before you build on it.
Drawback 4Not every system is connected
In practiceEach product is strong on its home turf: Muse in Meta's tools, dots in the apps connected to ChatGPT, Duet in email, calendar and bookkeeping. When we checked Muse, dots and Grok Bot, we found no ready connection to WordPress in any of them, and it isn't among the connectors Anthropic lists for Duet either. If your work sits in a website or in local software, check the connection before anything else.
Our opinion: a ready-made product is not inferior to an agent of your own. For personal work like email, calendar, documents and research, it may be entirely enough. An agent of your own justifies itself when it touches what brings money into the business.
Open-source operating software: what do you gain by choosing the harness yourself?
With an agent of your own you also choose the operating software, and you can choose open-source software. Open source means the code is visible to everyone, you install it on your own computer, and nobody rents it to you. One example is Hermes Agent by Nous Research, which is distributed under the MIT license. And there are others.
A ready-made product
- The model: the vendor chooses. You get the model the product runs.
- Customization: the connections, rules and channels the product offers.
- Memory and instructions: kept inside the product, with the vendor.
Open-source operating software
- The model: you choose, and you can switch. For example a strong model for writing a post, and a cheap model for a check that repeats every morning.
- Customization: skills you wrote, a connection to any software that has a way in, a schedule and a chat channel of your choice. For example local software that has no ready connection in the products.
- Memory and instructions: files on your computer. You can read them, back them up and move them to another computer.
This is how it looks in Hermes Agent, according to its documentation. It works with models from OpenAI, Anthropic and Google, through OpenRouter, or with a model that runs on your own computer, and you switch models with a single command. A scheduled task can be given a model of its own. You write to the agent on Telegram, WhatsApp, Slack or email. The memory, the skills and the conversation history are kept in one folder on the computer.
The price: open-source software comes with no support desk and no warranty. Someone has to install it, update it and secure it, and a mistake in permissions is yours. That is the work we do when we set up business automation with an AI agent. We usually build our clients’ agents on Hermes Agent, and choose the model that fits each business.
Who it suits: a business where the agent touches the website, advertising or customers, and anyone who cares about switching models without starting from the beginning. For personal work like email and calendar, start with a ready-made product.
Which way suits you? Five questions
Answer "yes" if the sentence is true for your business today. The result appears after the fifth question.
The quiz doesn't save or send your answers. And if you have no work that comes back every week, or nobody to approve, it's probably still too early for any of the ways.
How do you start with a ready-made product?
You start with one reading job, not by connecting the whole business. The common mistake is to connect everything on day one "so it has context." The opposite is better.
The first week with a ready-made product
- Check that it's open to youIn the app or the account, from your country, and without bypassing restrictions
- Connect one appCalendar or email, at read-only level
- Give it one jobFor example a morning summary of what's waiting for you
- Check for a weekAnd only then add a connection or permission to act
What you don't connect on day one:
- An account with access to money. The ad account, payment processing and bookkeeping wait until you've seen how it works.
- Customers' personal information. It is subject to the law and to the business's privacy policy. We expand on this in the guide to Muse and privacy.
- Permission to send or publish on its own. Everything that goes out starts with "ask first."
More on choosing permissions, for any kind of agent, in the article on AI agent security.
How to build an AI agent yourself: eight steps
You build an agent the way you bring on a new employee: a defined role, limited access, a supervised start, and only then more responsibility. You don't write the operating software from scratch: you take existing software, for example open-source operating software, install it on the cloud computer and configure it.
- Define 2–3 tasks. Each in one sentence: what comes in, what the agent does and what comes out. "When I send photos and a price, a product goes up as a draft," not "manage the store."
- Choose a model. OpenAI recommends starting with the strongest model to set a bar, and then checking whether a smaller, cheaper model still meets it. You test on your own examples.
- Connect software with minimum permission. A separate user for the agent in each system, with only the access the task needs. Before a change to the site there must be a backup and a way back.
- Write instructions and organize knowledge. OpenAI recommends building them from documents that already exist in the business. In a small business those are examples of work you liked, a tone, a price list and a "don't" list.
- Set approval rules. Publishing, budget, a message in the business's name and a significant change to the site wait for your approval. Deletion that can't be undone and payment details are out of bounds. You check that such an action really stops.
- Test on real jobs. Take jobs from recent weeks and compare the output with what was actually done. According to OpenAI, human involvement is especially important at the start, because that's where the edge cases surface.
- Turn on the log and tracking. When a mistake repeats, you fix the instruction or the permission, not just the output.
- Expand one task at a time. OpenAI also recommends one agent that gets tools gradually, before splitting into several agents.
An example: the eight steps in one business
Click through the steps. This is an illustration of a small online ceramics store, not a record of a real business.
Steps 1–2 · Define the task and choose a model
Every time a shipment arrives, I waste an evening adding products. That's what I want to hand over.
The task in one sentence: when you send photos and a price list, every item goes up as a draft product, with a description, price and category.
✓ Task defined · ✓ Two models tested on the same 10 examples
Step 3 · Connections and permissions
A separate user was created on the site for the agent, with access to products only. It has no access to plugins, users or settings. Before starting, we checked that there is a current backup.
✓ Read an existing product · ✓ Created one draft · ✗ No access to settings
Steps 4–5 · Instructions and approval rules
I'm attaching five descriptions I like. And no "luxurious," that's not us.
Noted. And the rules: I create a draft on my own. Publishing and a price change wait for your approval. Deleting a product is blocked.
On its own: drafts · With approval: publishing and price · Blocked: deletion
Step 6 · Testing on real jobs
We went over last week's shipment: 12 items. 10 passed with no corrections. On two I rounded a price, because the price list was handwritten.
The correction is saved for next time: an unclear price isn't rounded, it's flagged and you're asked.
! 2 corrections · ✓ Correction saved · ✓ Re-test passed
Steps 7–8 · The log, and only then expansion
In the log: 6 drafts created, 4 published with your approval, 2 waiting for a price check. After a few stable weeks a second task comes in: an Instagram post for every new product.
✓ Task 1 stable · ← Task 2 comes in
An illustration. The business, numbers and conversations in the example are made up.
When someone builds it for you: what to prepare, and what stays yours
You don't need to know how to code, but you do need to be there, especially at the start. A lot of the time goes to things that have nothing to do with technology: who holds the password to the ad account, what counts as a "good post," and where the site's backup is. Mark what you already have.
What to prepare before building an agent
Without the first three items there is nothing to build yet. The rest is usually settled on the first call.
You bring the tasks, the examples and the approvals. Whoever builds brings the computer or the setup of the ready-made product, the connections, the backup and the fixes when something changes.
What stays yours: the accounts, the instructions, the memory and the files. Write that into the agreement, so you can take them with you if you stop working together.
Red flags in a proposal to build an AI agent
Got a proposal? Go through it with these five questions. A single "no" is not a reason to walk away, but it is a reason to ask.
And three promises we would not accept: "one agent that manages the whole business from day one," "give it admin access, it's simpler," and "we build it and we're done." A good agent is built task by task, with minimum permission, and with someone responsible for it when the software around it changes.
For those who want to go deeper: what infrastructure do you run an agent on when it works for hours?
No single infrastructure is "the most cost-effective." It depends on how long each task runs, which software the agent touches, and how much control you need over permissions and information.
A task that lasts hours, such as going through every page of the site, needs a place where the work is saved, and a way to recover from a failure or stop for approval without losing what's been done. Anyone looking for infrastructure for long-running agent work is choosing between a managed platform, a dedicated cloud computer and a ready-made product.
A managed agent platform. OpenAI's Agents API opened in public beta on September 10, 2026. According to the documentation, OpenAI manages the sessions (the work that is saved between one stage and the next), the coordination and the recovery from failures. The developer supplies the tools and chooses where the agent works: with no environment, in an environment hosted by OpenAI, or in their own. On September 29 computer use was added, an agent that works in a browser hosted by OpenAI, and according to the documentation it currently requires GPT-6 Astra.
Your own agent on a dedicated cloud computer. The agent runs on a computer allocated to one business, and calls the model through the vendor's official connection (API). You, or whoever builds with you, set the users, the approval rules, the log and the backups. The trade-off is that someone has to maintain all of that.
Three kinds of infrastructure for long-running work
A managed platform (Agents API)
- Cost: by use: model, tools and environment run time
- Control: high in code, within the limits of a beta API
- Information: according to the documentation, currently US-only storage and no Zero Data Retention
- Maintenance: with the developer: code that may change during the beta
- Fits when: you have a developer, and tasks run for hours in an isolated environment
A dedicated cloud computer
- Cost: a computer at a fixed monthly fee, the model by use, and maintenance
- Control: full: users, permissions, approvals and log
- Information: on your computer. Calls to the model go to the vendor
- Maintenance: with whoever manages the computer
- Fits when: the agent touches the website, advertising and the business's software
A ready-made product (dots, Muse, Grok Bot, Duet)
- Cost: a subscription, or free use with limits
- Control: within the rules the product offers
- Information: in the account with the vendor
- Maintenance: almost none
- Fits when: the tasks are personal and the apps are supported
What makes long work expensive? Mainly the history the task drags along and processes again and again, because you pay for the model by the amount of text it reads and writes. So it's worth breaking long work into stages with save points, and checking whether a smaller, cheaper model is enough for it. Current prices are on each vendor's pricing page.
Our opinion: for a small business whose agent touches a WordPress site, Google Ads and Meta, a dedicated computer per business is today the safer choice. The Agents API is interesting, and it also offers an option for an environment in your own infrastructure, but according to the documentation even that doesn't make the service eligible for Zero Data Retention.
Tools like n8n have done good work with fixed sequences. A long task is full of small decisions, and there an agent fits better. We expand on this in the comparison of n8n and an AI agent.
How much does it cost to build an AI agent?
The price follows from the way you chose. With a ready-made product you pay a subscription to the vendor. With an agent of your own there are three components: a one-time setup, external services such as the model and the cloud computer, and ongoing support. With us, as a ballpark: a basic setup is about ₪2,000, each connection adds ₪500 to ₪2,000, and monthly support starts at ₪1,000. An exact price is given in writing after an intro call. You can build an estimate yourself with the price calculator.
When you compare proposals, compare part by part and not just the bottom line. An unusually cheap proposal usually builds fewer than the ten parts.
Where does Platin Digital come in?
We usually build the agent on a separate cloud computer used only by your business, with a separate user and minimum permission for each system that's connected. Publishing, budget and messages to customers always wait for your approval, and you write to the agent on WhatsApp, Telegram or email.
We also set up and configure agents on the ready-made products, when that's what suits the business. If Muse, a dot, Grok Bot or Claude Duet will be enough for you, we'll say so, and help connect it and write the rules. The model is chosen by the work: from OpenAI, Anthropic, Google and others. Platin Digital is an OpenAI Select Partner.
What the agent does once it exists, by type of work:
- Business automation with an AI agent and an AI customer service and sales agent
- An AI personal assistant for your business for email, calendar and tasks
- Website and store management with an AI agent and SEO with an AI agent
- Ad management with an AI agent and social media management with an AI agent
And what it looks like in your trade: AI agent for lawyers, AI agent for real estate and AI agent for ecommerce. And if you want to see a whole day, read what an AI agent does on a normal day.
Frequently asked questions
What does it mean to build an AI agent?
Building an AI agent is connecting several parts into one system: a model that understands and writes, operating software that carries out actions for it, a computer it runs on, connections to the business's software, memory, instructions, a chat channel, permissions and a log. You can get all of this as a ready-made product, build it on a cloud computer of your own, or ask someone to set it up.
Can I build an AI agent myself without knowing how to code?
Yes, in two ways. With a ready-made product like Muse, dots, Grok Bot or Claude Duet you write no code at all: you connect apps and write rules in plain language. An agent on a cloud computer of your own doesn't require writing software either, but it does require installing, setting permissions and maintaining. If the agent touches the website, advertising or money and you have no technical background, it's better to have someone set it up with you.
Which is better: Muse, dots, Grok Bot or Claude Duet, or an agent on my own server?
It depends on what the agent needs to do. A ready-made product is easier to start and connect, and suits personal work like email, calendar and research. An agent on your own server gives control over the permissions, the computer and the model, and fits when it works in the website, in advertising or with customers. With a ready-made product the permissions, the computer and the model are the vendor's, and you're locked into one system. Your own server has to be set up and maintained.
What is an open-source harness, and why does it matter?
A harness is the operating software: the software that runs the model and carries out actions for it. When it is open source, its code is visible and you install it on your own computer. One example is Hermes Agent by Nous Research. It matters because this way you choose the model and can switch it, you add skills and connections of your own, and the memory stays with you as files. The price: someone has to install it, update it and secure it.
How do you build an AI agent?
You define 2–3 recurring tasks, choose a model by testing on real examples, and connect the software with a separate user and minimum permission. Then you write instructions, set what needs approval, test on jobs from recent weeks, turn on a log, and only then expand.
How long does it take to build an AI agent?
A ready-made product you can start trying in minutes, if it's open to your account. An agent on a computer of your own usually takes weeks, including testing on real jobs, and it depends on the number of tasks and systems and on how organized access to them is. Anyone who promises an agent that runs the whole business within a day is probably skipping some of the parts.
Which models and tools does Platin Digital build agents with?
Platin Digital is an OpenAI Select Partner. We build agents with the model and tools that fit each business, from OpenAI, Anthropic, Google and others. Usually the agent runs on a separate cloud computer for each business, with a separate user and minimum permission in each system and approval rules defined in advance. When a ready-made product is a better fit, we set it up and configure it.
Sources
- https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
- https://developers.openai.com/api/docs/guides/function-calling
- https://developers.openai.com/api/docs/guides/agents/guardrails-approvals
- https://developers.openai.com/api/docs/guides/your-data
- https://developers.openai.com/api/docs/guides/agents-api/overview
- https://developers.openai.com/api/docs/guides/agents-api/sessions
- https://developers.openai.com/api/docs/guides/agents-api/environments/self-hosted
- https://developers.openai.com/api/docs/guides/agents-api/tools/computer-use
- https://developers.openai.com/api/docs/changelog
- https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
- https://about.fb.com/news/2026/09/introducing-muse-small-business/
- https://www.meta.com/help/subscriptions/1021145227643680/
- https://learn.chatgpt.com/docs/dots
- https://learn.chatgpt.com/docs/dots/controls
- https://learn.chatgpt.com/docs/dots/computers-and-apps
- https://docs.x.ai/grok-bot/faq
- https://docs.x.ai/grok-bot/security
- https://docs.x.ai/grok-bot/skills-routines-and-automations
- https://cursor.com/help/grok-bot/plans
- https://claude.com/resources/articles/claude-duet
- https://claude.com/product/claude-duet
- https://support.claude.com/en/articles/17631495-get-started-with-claude-duet
- https://nousresearch.com/
- https://github.com/NousResearch/hermes-agent
- https://hermes-agent.nousresearch.com/docs/
- https://hermes-agent.nousresearch.com/docs/integrations/providers
- https://hermes-agent.nousresearch.com/docs/user-guide/messaging/
- https://hermes-agent.nousresearch.com/docs/user-guide/features/memory
- https://hermes-agent.nousresearch.com/docs/user-guide/features/skills
- https://hermes-agent.nousresearch.com/docs/user-guide/features/cron
- https://x.ai/
