> Source: https://builder.ema.ai/v2/builder-guides/orchestrator-ai-employee > Title: Creating an Orchestrator AI Employee # Creating an Orchestrator AI Employee This guide builds an **Orchestrator AI Employee**: an AI Employee that takes a request in chat and coordinates other AI Employees, knowledge bases, and tools to complete it. There is no workflow to wire. You configure it on one tab, publish, and chat. If you have not met orchestrators yet, read [Orchestrator AI Employees](/builder/v2/core-concepts/orchestrators) first. > [TIP] > **Prefer to describe it?** [Autopilot](/builder/v2/autopilot) can create and configure an orchestrator from a plain-language description, and once it is published, run tasks against it to test it. The steps below cover the same settings, so you can review and adjust whatever Autopilot builds. ## What you will build The example throughout is **Employee Help**, an orchestrator for an internal help desk. It answers general questions from a company handbook, hands time-off questions to an existing **PTO Assistant** AI Employee, and hands hardware and access requests to an **IT Helpdesk** AI Employee. Before you start, make sure the AI Employees you want to connect are **published**. An orchestrator cannot delegate to a draft. ## Step 1: Create the orchestrator 1. Open the **AI Employees** page and click **Build an AI Employee**, then choose to start from scratch. 2. Enter a **Name** (for example, `Employee Help`) and a **Description**. Both are required. 3. Optionally pick a **Category** to group it on the AI Employees page. 4. Under **Interaction type**, choose **Orchestrator** ("Coordinates other AI Employees, knowledge, and tools to handle multi-step requests"). 5. Click **Create**. You land on the orchestrator's page. Orchestrators have their own set of tabs: **Chat**, **Configuration**, **Training**, **Permissions**, **Audit**, **Metrics**, and **Activity**. **Training** is where you [evaluate the orchestrator](/builder/v2/testing-operations/evaluating-orchestrators) against test datasets. There is no workflow editor, and the **Edit workflow** action does not appear in the header menu. Open the **Configuration** tab. The header shows the orchestrator's name, its current version and status, and the **Save draft** and **Publish** buttons. Everything below the header is a setting you can change. ## Step 2: Write the Instructions The **Instructions** card is where you tell the orchestrator who it is and how it should help. Markdown is supported. Good Instructions cover: - **Role and scope.** What it helps with, and what it should decline or redirect. - **When to delegate.** Which kinds of requests go to which connected AI Employee, by name. - **When to answer itself.** For example, general questions it can answer from its knowledge bases. - **Tone and format.** How it should speak, how long replies should be, when to confirm before acting. For Employee Help: ```text You are Employee Help, the first stop for employee questions at Acme. - Time off, PTO balances, and leave policy: hand the request to the PTO Assistant and relay its answer. - Laptops, software access, passwords, and IT incidents: hand the request to the IT Helpdesk. - General questions about company policy and benefits: answer from the employee handbook. Cite the document you used. - If a request spans several areas, handle each part with the right AI Employee and give one combined answer. Keep replies short and friendly. Confirm with the person before submitting anything on their behalf. ``` Refer to connected AI Employees by their names. The orchestrator sees those names and uses them to decide where to send work. The general advice in [Writing effective Instructions](/builder/v2/builder-guides/writing-instructions) applies here too. ## Step 3: Connect AI Employees The **Connected AI Employees** card sets which AI Employees the orchestrator may delegate to. - **Pick specific AI Employees.** Search and select each one. For Employee Help, select **PTO Assistant** and **IT Helpdesk**. You can connect Chat AI Employees, Dashboard AI Employees, and other orchestrators. Delegating to a Voice AI Employee is not supported yet; support is planned for a future release. - **Or connect all of them.** Turn on **All AI Employees** to let the orchestrator use every AI Employee in your workspace, including ones added later. Turn it off again to go back to a specific list. You can only connect AI Employees you have access to. If the orchestrator later builds an AI Employee itself (see Step 7), that AI Employee is connected automatically and appears under a **Created by this orchestrator** heading in the picker. > [TIP] > Start with a short, specific list. Every connected AI Employee is one more choice the orchestrator has to make on each request. Add more once the core routing works. ## Step 4: Add knowledge bases The **Knowledge bases** card works the same way as on any other AI Employee: create a knowledge base, upload documents or connect a source, and the orchestrator can search it when it answers a question itself. For Employee Help, add the employee handbook so general policy questions get answered directly instead of being delegated. If you use tagging, the **Tags** card lets you set up taxonomies and auto-tagging for the orchestrator's sources. See [Setting up a knowledge base](/builder/v2/builder-guides/knowledge-base-setup) for the full walkthrough. ## Step 5: Upload skills (optional) A **skill** is a Markdown file that gives the orchestrator reusable know-how for a specific kind of task: a step-by-step process, a checklist, house rules for a recurring request. The orchestrator draws on a skill automatically when the conversation calls for it, so you do not have to repeat the process in your Instructions. 1. On the **Skills** card, click **Upload skill**. 2. Choose one or more `.md` files. Only Markdown files are accepted. 3. Each uploaded skill appears in the list. Click the **X** next to a skill to remove it. A skill file is plain Markdown. Give it a clear title, say when it applies, and write the guidance the way you would hand it to a new teammate: ```text # Scheduling an interview Use this when someone asks to set up a candidate interview. 1. Confirm the candidate's time zone and availability window. 2. Check the interviewer's calendar for a 45-minute slot. 3. ... ``` Skills stay private to this orchestrator and go live when you publish. ## Step 6: Choose built-in tools The **Built-in tools** card lists general abilities you can turn on or off. Core abilities such as searching knowledge bases and delegating to connected AI Employees are always available and are not listed here. Tool What it does **Create files** Write and edit files while it works, such as drafting a document to hand back. **Web access** Search the web and read pages. **Memory** Remember a person's preferences and learnings across chats. **Interactive widgets** Show rich, interactive responses in chat, such as tables and charts. All four are on by default. Turn off anything the orchestrator should not do. For an internal help desk, you might turn off **Web access** so answers come only from company sources. ## Step 7: Decide on management abilities The **Management abilities** card controls whether the orchestrator can **create and change** platform resources. Reading and using connected resources is always allowed; these switches govern building and editing them. Ability What it allows **Build & edit AI Employees** Create and edit AI Employees and their workflows from a conversation. **Manage knowledge bases** Create knowledge bases and manage their documents and syncs. **Manage integrations** Build, connect, and manage integrations. All three are **off** by default. Leave them off for an orchestrator that serves end users, like Employee Help. Turn one on only when the orchestrator's job is to build or administer things, for example an orchestrator your team uses to spin up new AI Employees. > [WARNING] > Management abilities let the orchestrator make real changes in your workspace on behalf of whoever is chatting with it. Combine them with a tight [permissions list](#step-9-set-permissions-and-publish). Under **Capabilities**, the **Sub-agents** switch lets the orchestrator split a big request into parts and work on them in parallel. It is on by default. Turn it off if you want every request handled one step at a time. ## Step 8: Set limits Under **Limits**, **Concurrent chats per person** caps how many chats one person can have running with this orchestrator at the same time. The default is 3. Enter `0` for no limit. When someone hits the limit, they are asked to end or close one of their running chats before starting another. ## Step 9: Set permissions and publish 1. Click **Save draft**. The header status changes from **Unsaved changes** to **Draft saved, not published**. 2. Click **Publish**. The status changes to **Published, up to date** and the version badge increments. The first publish also enables the orchestrator, so people can start chatting right away. 3. On the **Permissions** tab, choose who can use the orchestrator. Permissions work the same as for any other AI Employee; see [Roles and access](/builder/v2/core-concepts/ai-employees#roles-and-access). **Publish** is only available when there is a saved draft that has not been published yet. If you leave the Configuration tab with unsaved edits, you are asked whether to save or discard them. ## Step 10: Test it in chat Open the **Chat** tab. The orchestrator greets you ("Hi, I'm Employee Help. How can I help you today?") and you can start typing. Try each kind of request your Instructions cover, and at least one that spans two areas: - "How many PTO days do I have left?" should go to the PTO Assistant. - "My laptop won't turn on." should go to the IT Helpdesk. - "What's the parental leave policy, and can I book two weeks in March?" should answer the policy part from the handbook and hand the booking to the PTO Assistant, then give one combined answer. As the orchestrator works, the chat shows which AI Employee it handed each part to and what came back. To see every individual step instead of the summary, open the menu in the page header and switch the transcript from **Narrative** to **Developer** view. Switch back the same way. Each chat appears in the session list beside the conversation, so you can start a new chat, return to an earlier one, or compare how two conversations went. You can also ask [Autopilot](/builder/v2/autopilot) to send test requests to the published orchestrator and report back what it did, which is a quick way to check routing without opening the Chat tab yourself. Chat shows you how the orchestrator handles the requests you think to try. To test it against dozens of realistic requests at once and see which ones fail, [evaluate it](/builder/v2/testing-operations/evaluating-orchestrators) on the **Training** tab. If the Chat tab tells you to publish first, go back to Configuration and click **Publish**. If it says the orchestrator is disabled, turn it on with the switch in the page header. > [TIP] > When a test goes wrong, the fix is usually in the Instructions: name the AI Employee explicitly, describe the kinds of requests it owns, and say what to do when a request does not fit anywhere. Save, publish, and try again. ## Step 11: Iterate with version history Every publish creates a numbered, immutable version. Open **Version history** from the header menu to see them; the live one is marked **Current**. To go back to an earlier version: 1. Click **Restore** on that version. 2. Confirm **Restore to draft**. The version's configuration is loaded into your draft. Nothing changes for people using the orchestrator yet. 3. Review the draft on the Configuration tab, then **Publish** to make it live. ## Step 12: Manage the orchestrator The header menu offers the same management actions as other AI Employees, with a few orchestrator-specific notes: - **Clone** creates a copy with its own Instructions, settings, and skills, as a draft you publish to activate. Connected AI Employees are shared with the original, not copied. - **Export** and **Import** carry the orchestrator's configuration and its list of connected AI Employees as a bundle, so you can move it between workspaces. See [AI Employee export and import](/builder/v2/administration/ai-employee-export-import). - **Save as template** lets others in your workspace create their own orchestrator from this one. Only a name and category are needed. - **Enable / disable** in the header stops or resumes new chats. The published configuration is kept, so re-enabling does not need a new publish. - **Rename** and **Delete** work as for any AI Employee. ## Reviewing how it is used - The **Audit** tab lists every conversation people have had with the orchestrator, with its status, number of turns, which AI Employees were involved, and whether it paused for a person. Open a session to read the transcript and see how long each turn spent on each step. - The **Metrics** tab shows usage over time. - On a connected AI Employee, runs the orchestrator started are marked **Triggered by** the orchestrator's name, so you can follow a request from either side. ## Where to go next - [Orchestrator AI Employees](/builder/v2/core-concepts/orchestrators), the concept page, including [best practices](/builder/v2/core-concepts/orchestrators#best-practices) such as one orchestrator per group. - [Writing effective Instructions](/builder/v2/builder-guides/writing-instructions) for more on Instructions that produce reliable routing. - [Evaluating Orchestrator AI Employees](/builder/v2/testing-operations/evaluating-orchestrators) to test the orchestrator against a dataset of simulated conversations after each change. - [Autopilot](/builder/v2/autopilot) to build and edit orchestrators by describing what you want.