How to Build AI Employees: From One Agent to a Team, No Terminal Needed

The idea of an "AI employee" has caught on, and so have guides to building one. Most of them run Claude Code in a terminal. If you don't write code, that first step is the hardest part.

This guide builds a first AI employee without a terminal, turns it into a team from the second one on, and hands it weekly work — all in one desktop app, Orkas. It includes a job-description template you can copy, and how other tools compare.

The short version

1. Start with one job. Pick work that repeats every week, that you could explain to a person, and whose result you can check in five minutes.
2. Write one job description. Decide the role, inputs, outputs, rules, pass criteria and when to stop.
3. Widen the scope as you check. From the second AI employee on, chain them into a relay, and keep the final check with a person.

What is an AI employee? Role, memory and continuity

In this guide, an AI employee is an AI that handles a set job, the same way, over and over.

When you ask a chatbot for help, you explain every time what you want, in what form and what to watch out for. An AI employee keeps that as a job description and remembers what has been decided so far. From the second run on, "please do this week's batch" is enough.

People often mix it up with "AI agent". An AI agent is the mechanism; an AI employee is a way of using one. In Orkas, you build an AI employee as an agent.

How to choose the first job

Your first AI employee is least likely to fail on work that meets these four conditions.

• It repeats every week. A one-off job isn't worth writing a job description for.
• You could explain it to a person. If you can't explain it to a person, you can't hand it to an AI either.
• Mistakes can be undone. Pick work a person can stop before it leaves the company.
• The result can be checked in five minutes. If checking takes an hour, it's faster to do it yourself.

For a small company, good first jobs look like this.

Don't hand over approvals involving money, the final send to a customer, or decisions about personal data — at least not at first. Let the AI employee draft; people decide.

How to write the job description (template)

The job description decides most of how good an AI employee is. Write one per AI employee, covering these eight points.

1. Name and role: what it owns, and what "done" means.
2. Inputs: what it may read, and what it must not.
3. Outputs: what it produces, in what format, and where it goes.
4. Rules: tone, words to avoid, facts it must not invent.
5. Check criteria: what passes.
6. When to stop: when it should hand the work back to a person.
7. When it runs: when it starts work.
8. Review: where corrections get written down.

Here is a filled-in example for an online shop's social media writer. Replace the details with your own.

AI employee job description

1. Name and role
- Name: SocialPosts
- Role: every week, write three Instagram post ideas from the new-product information
- Goal: the shop owner can check them in five minutes and schedule them

2. Inputs
- Sources: "This week's new products" and the brand guide in the Library
- Examples: the ten past posts that got the best response
- Must not read: customer lists, sales data, any file with personal data

3. Outputs
- Deliverable: three posts + hashtags + image directions
- Format: a heading for each post; body under 200 characters
- Where: save to the Library as "date social post ideas"

4. Rules
- Tone: friendly and polite; at most two emojis per post
- Words to avoid: "cheapest", "guaranteed", "No. 1 in the industry"
- Facts: use only the prices, stock and release dates given in the sources

5. Check criteria
- Product names and prices match the sources
- Each post is under 200 characters
- All three posts open differently

6. When to stop
- A product has no price in the sources
- Anything about complaints, returns or discounts
- When unsure, don't write it: note the question and stop

7. When it runs
- Every Monday at 9 a.m., and whenever the shop owner asks

8. Review
- Add each human correction as one line under "4. Rules" or "5. Check criteria"
- If the same correction comes up three times, rewrite this job description

If you can't fill in every point, the job isn't ready for an AI yet. Write down how a person does it first.

Five steps to put your first AI employee to work in Orkas (no terminal)

Step 1: Install Orkas and choose your AI

Orkas is a desktop app for macOS and Windows. When you install it and sign in, you get 500 credits and can use Orkas's official models right away. There are two other ways to use AI:

• Use your own API key. Under Models & credentials → Add model, add a key from OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter and others.
• Connect Claude Code or Codex on your computer. Add them as external agents from Connect agents. Each tool must be installed and signed in with your own account.

Step 2: Create a project and gather the material in the Library

Create a project for each body of work. Put everything the AI employee reads — product information, the brand guide — in one place with Add to Library.

Write the rules the whole company follows in the project's Goals & rules. You have up to 4,000 characters, and every task in the project reads them first.

Step 3: Create the agent and give it the job description

Click Create agent to open New agent. Enter a Name and a Summary (one line on what it does), and in the editing chat that follows, the AI writes its Workflow — the working procedure. Paste in your job description there and ask it to "shape the workflow and delivery standards from this." The pass criteria go into Delivery standards (up to five).

You can also skip the form and ask Commander, the AI that directs your other agents in Orkas, in a chat: "I want someone who drafts our social posts every week."

Use letters, numbers and Chinese or Japanese characters for the name, with no spaces (for example, SocialPosts). Names are used to @-mention agents, so spaces and most symbols aren't allowed.

Step 4: Try it on real work

Start a new task in the project and ask your agent for this week's batch. The first result is usually slightly off in shape. Adjust the output format — headings, length — until you can check it in five minutes.

Step 5: Write the corrections back

Don't leave corrections in the chat; write them back. Anything specific to that agent goes into its workflow or Delivery standards in the editor; rules for the whole project go into Goals & rules. When the same correction comes up three times, it's time to revisit the job description itself.

The screens are covered step by step in the agents chapter of the docs.

Memory and handover: never explain the same thing twice

Memory in Orkas has three layers.

• Agent memory: used only by that AI employee.
• Project memory: what has been decided in that project.
• Global shared memory: long-term facts used across conversations and agents.

The AI saves facts that look useful in the long run on its own, without being asked. Only Commander writes project memory and global shared memory; specialist agents write only their own. You can add, edit, import and export entries at any time on the Memory screen. Each layer holds up to 16 entries, so keep only what matters.

• Keep: decisions, corrections that keep coming up, each client's preferences.
• Don't keep: personal data, passwords, one-off instructions.

From the second one on, make it a team: Commander runs the relay

Once the first AI employee is stable, add a second and a third for the steps before and after it. For an online shop's social media, the split might look like this:

1. Researcher: pulls the selling points from the new-product material and past reactions.
2. Social writer: writes three post ideas from those points.
3. Checker: checks them against the criteria and sends them back if they fail.
4. Shop owner (a person): gives the final check and schedules the posts.

Write one job description per role and create each as an agent. There are two ways to connect them:

• Let Commander run it. Ask for "posts for this week's three new products", and Commander splits the work by role, takes each result and passes it to the next. Steps that don't depend on each other can run at the same time.
• Set the order yourself. In a chat, @-mention several agents and choose Run in order.

To make it weekly, use the project's To-do list. Choose the agent for a to-do and select Process now; when it's done, it moves to Awaiting confirmation, and a person checks it and sets it to Done. Turn on the to-do list's Auto switch and Commander advances ready to-dos one at a time (up to 20 a day).

How Commander works is explained on the Commander page.

Daily and weekly work: automations and their limits

Register work that should run at set times under Auto in the sidebar. Frequency can be Once, Every x hours, Daily, Weekly or Monthly. After setting one up, try it once with Run now.

Know these limits first:

• Automations run only while Orkas is open.
• Runs missed while the app was closed are not made up later. That includes one-time automations.
• They run only on the computer where you set them up.
• On Windows, closing the window also quits the app.

The safest choice is to schedule them for hours when your work computer is on. If the work has to run while your computer is off, look at tools that run in the cloud (see the comparison below).

What people still do: final checks, decisions and confidential data

Even with AI employees, keep these three jobs with people.

• Checks before anything goes out. Posts, emails and page updates go out after a person has looked.
• Money and promises. People decide discounts, refunds and delivery dates.
• Decisions about personal data. People decide what information may be used.

Be careful with confidential information, too. Orkas requires you to sign in, and AI processing runs on cloud models. Choose what you give your AI employees according to your company's rules, and leave customers' personal data out. API keys and connector sign-ins belong to each person who uses them.

You can also run this with human colleagues. Since Orkas 1.8, you can share a project with other people. Free accounts can share up to 3 projects with up to 3 members each, including the owner; members can join with a free account. When you assign a to-do to a member, their AI handles it on their computer. What's shared is the goals and rules, memory, Library, to-dos and the agents you share; conversations are not shared.

How to build one in other tools: Claude Code, Codex, ChatGPT, Gemini, Copilot

You can build AI employees outside Orkas too. Here is the main approach in each tool as of October 2026. Check each vendor's official pages for current details.

If you're comfortable with development and your work lives in a repository, Claude Code or Codex used directly is the most powerful option. If something must run on a schedule while your computer is off, a cloud setup such as Claude Code's Routines fits better.

If your work isn't programming and you want several roles to hand work to each other, with material, to-dos and automations in one place, Orkas fits. You can also connect Claude Code or Codex on your computer as Orkas agents; see how to use Claude Code and Codex together.

Sources: Scheduled tasks in ChatGPT, Scheduled actions in Gemini, Agent Builder in Microsoft 365 Copilot, Routines in Claude Code

What it costs

Orkas pricing as of October 2026. Check the pricing page for current prices.

• Free: free to download, with 500 credits when you sign in to the desktop app.
• Lite: $11.99 a month, with 5,000 credits each month.
• Pro: $19.99 a month, with 10,000 credits each month.
• Max: $79.99 a month, with 50,000 credits each month.

With your own API key, you pay the key's provider for AI usage. Connected Claude Code or Codex runs within that tool's own plan.

Common mistakes and how to avoid them

• Leaving it all to the one tech-savvy employee. The person who knows the job best writes the job description; the tech-savvy person helps.
• Starting without written rules. Corrections given out loud are lost. Add each one to the job description.
• Automating everything at once. For the first few weeks, a person checks every result. Widen the scope after you've seen the stop conditions work.
• Making the role too broad. Not "all of marketing" but "three post ideas a week": one deliverable.
• No pass criteria. Write down what counts as done in the check criteria.
• Mixing in confidential data. Keep customer lists and passwords out of both the material and the memory.

FAQ

Do I need to know how to program?

No. In Orkas, creating agents, adding material and setting up automations all happen on screen. You don't use a terminal.

Can I build one for free?

Yes. Orkas is free to download, and you get 500 credits when you sign in to the desktop app. You can also use your own API key.

How is this different from projects in ChatGPT or Claude?

Both bundle instructions and files. On top of that, Orkas lets several agents hand work to each other, runs the work weekly with to-dos and automations, and connects Claude Code or Codex on your computer — the parts that make it a team, in one app.

Does it keep working when my computer is off?

No. Orkas automations run only while the app is open, and runs missed while it was closed are not made up later.

Does it work in languages other than English?

Yes. The app is available in many languages, including Japanese. The names of the built-in agents, such as ContentWriter, stay in English.

Is it safe to give it company documents?

Decide according to your company's rules. Orkas requires you to sign in, and AI processing runs on cloud models. Don't give it customers' personal data or passwords.

Wrapping up

Building AI employees starts less with picking a tool than with one job and one job description. Once the first is stable, add roles before and after it and let Commander run the relay. Keep checks and decisions with people. In that order, you can run it as a team without programming.

Hand work to a team of AI specialists and get team-sized results on your own. Orkas is free to download.