--- title: Create and manage assistants description: Create a private or shared assistant in Agent Studio, write its instructions, pick a model, try and test it, then publish it or share it with colleagues. url: https://agentlab.cresclab.com/docs/en/studio/assistants language: en updated: 2026-11-06 --- # Create and manage assistants Create a private or shared assistant in Agent Studio, write its instructions, pick a model, try and test it, then publish it or share it with colleagues. ## Private and shared assistants The **Assistants** page has three tabs: **All**, **Shared** and **Private**. The two kinds of assistant differ in who can use them: | | Private assistant | Shared assistant | | ------------------ | ------------------------------------------------------------------------------------- | ------------------------------------------------------------------------ | | Who can use it | Only you. Other members can't see that it exists, and admins can't read its contents. | Once published, the whole workspace or the teams you choose. | | When changes apply | Your next message uses what you saved. There is no publish step. | Changes are saved as a draft and apply only when someone publishes them. | | Channels | Not available. | Can be connected to Slack or LINE once published. | | How many | 20 per member. | 100 per workspace. Archived ones don't count. | You can later share a private assistant so it becomes a shared one ([Share a private assistant](https://agentlab.cresclab.com/docs/en/studio/assistants.md#share)). A shared assistant can't go back to private. ## Create an assistant 1. **Open the dialog** On the **Assistants** page, select **Create an assistant**. 2. **Choose who can use it** Under **Who can use it?**, choose **Only me (private)** or **The whole workspace (shared)**, then select **Next**. If you opened the dialog from the **Private** or **Shared** tab, this is already chosen. 3. **Choose a starting point** Select **Start from blank**, or pick a template under **Or start from a template** and select **Use template**. A template fills in the job, audience and hard rules, and comes with two test questions. The assistant's page opens as soon as it's created. A blank shared assistant is called **Untitled assistant** and a blank private one **Untitled private assistant**; you set its name and everything else on its page. > [!NOTE] A new shared assistant starts as a draft > Until it's published, only the person who created it can see it. Colleagues won't find it in their list. While a workspace has no assistants of its own, the **All** and **Shared** tabs show **Start from an example**. Selecting **Start from this** also creates an assistant with test questions already written. ## What's on an assistant's page A shared assistant's page has eight sections, with **Configuration summary**, **Before publishing**, **Instructions preview** and **Try it** in the side column: | Section | What it holds | | --------------------------- | --------------------------------------------------------------------------------------------------------- | | 1. The basics | **Assistant settings**, **Model**, **Visibility** | | 2. Skills | Skills attached from the skill library | | 3. Data sources | **Knowledge base** documents and **Q&A** | | 4. Channels | A link to **Configure channels** | | 5. Actions (advanced) | Calls to your own APIs, and **Connected systems** | | 6. Rules | The workspace rules that govern this assistant | | 7. Test your assistant | Test questions, test runs and publishing | | 8. Advanced and integration | **Instructions (advanced) — the system prompt**, **Version history**, **API keys**, **Archive assistant** | A private assistant's page is shorter: **What it does**, **Files it reads**, **Your connectors** and **Share or deploy**. The model, skills, test questions, version history and the rest are under **Other settings**. The side column offers **Talk to it** instead. ## Instructions An assistant answers by following its instructions. You don't have to write them yourself: answer a few questions in **Assistant settings** in your own words, and Agent Lab composes the instructions from your answers. | Field | What it's for | | ---------------------------- | ----------------------------------------------------------------------------- | | Assistant name | The name the assistant introduces itself with. Required. | | What should it help them do? | The scope of its job. | | Shape of what it hands you | Optional, for example "one slide per section: a headline plus three bullets". | **More settings** opens the rest: | Field | What it's for | | ----------------------------------------- | --------------------------------------------------------------------------------------------------------- | | Who is it for? | Who it serves. | | Pick a speaking style | **Professional and precise**, **Warm and friendly** or **Short and to the point**. | | Anything else about the style? (optional) | For example, always address guests politely. | | What must it never do or say? | Hard rules. The assistant refuses rather than break them. | | How it should work | Whether it checks with you first or hands over a draft straight away. | | When it does not have the facts | **Never state anything it cannot back up**, or **May propose, but must label proposals and assumptions**. | The **Instructions preview** in the side column updates as you type and shows exactly what the assistant will be told. On a shared assistant, **Save as a new draft** creates a new draft version, and nothing changes for anyone until it's published. On a private assistant, **Save** applies to your next message. To edit the full text of the instructions, open **Instructions (advanced) — the system prompt**, select **Edit instructions**, and save with **Save as a new draft** (**Save** on a private assistant). > [!WARNING] Saving Assistant settings again replaces manual edits > Saving **Assistant settings** recomposes the instructions from your answers, so anything you edited by hand in the advanced instructions isn't carried into the new draft. The earlier version stays in **Version history**, where you can restore it. ## Model **Model** sets which AI model the assistant uses. Unless you need a particular one, keep **Use the workspace default**; the default is set in [AI settings](https://agentlab.cresclab.com/docs/en/studio/ai-settings.md). The list may also offer **Capability** choices: pick one and the workspace's configuration chooses a suitable model. Pick a specific model and the assistant stays on that model. Models your plan doesn't include are marked **(not in your plan)** and can't be selected. If the assistant's current model is no longer in your plan, its replies are refused and the page tells you to pick another. Changing the model takes effect immediately, without publishing. ## Skills, knowledge and tools All of these are optional. An assistant can answer from its instructions alone. - **Skills** attach your team's know-how snippets to the assistant. See [Skills](https://agentlab.cresclab.com/docs/en/studio/skills.md). - **Knowledge base and Q&A** let the assistant quote documents, or always see answers someone wrote. See [Knowledge](https://agentlab.cresclab.com/docs/en/studio/knowledge.md). - **Actions and connectors** let the assistant call your APIs or reach outside systems. See [Integrations](https://agentlab.cresclab.com/docs/en/studio/integrations.md). ## Try and test it ### Try it On a shared assistant, open **Try it** in the side column, type in **Ask it something** and select **Try as a trial**. A trial talks to the saved draft and is labeled as a trial. Once the assistant is published, **Start a real conversation** opens a conversation with the live version in your **Conversations**. On a private assistant, select **Talk to it**, type a message and select **Start a conversation**. Every message uses whatever is saved at that moment. > [!NOTE] Conversations use saved settings > Trials, test runs and publishing ignore changes you haven't saved. If you see **Unsaved changes**, save first. ### Test questions Test questions check that the assistant gets the important things right. Each one has a **Short name**, **How would someone ask it?** and **What must a good answer include?**. You can add them three ways: - Fill in the fields under **Test questions** and select **Add question**. - In a conversation, open **Message actions** on a reply and choose **Save as a test question**. - Under **Import questions from a spreadsheet**, upload a CSV whose first row is `title`, `question`, `expected`, with 1 to 100 questions. Every row is checked, and nothing is imported until all of them pass. An assistant holds up to 100 test questions. You can also group, pause or delete them, or export them with **Download as CSV**. ### Run the tests In **Test your assistant**, select **Run the tests**. Every question is put to the saved draft, and AI marks each answer **Passed** or **Failed** against what you said a good answer must include. Someone with publishing rights can overturn a verdict they disagree with. If the assistant already has a live version, the results are compared with it: **Improved**, **Got worse**, **Still passing** or **Still failing**. **Past test runs** lists each run with its pass rate and token usage. > [!NOTE] Tests spend AI budget > A test run uses tokens the same way a conversation does. When the workspace's budget is used up, the run doesn't start. ## Publish a shared assistant Colleagues can use a shared assistant only after it's published. 1. **Save your settings** Make sure the header doesn't show **Unsaved changes**. 2. **Go to publishing** Select **Review and publish** in the header or side column. It takes you to **Test your assistant**. 3. **Test it (recommended)** Select **Run the tests** and read the results. 4. **Publish** Select **Publish this draft**. When you see **Your assistant is live**, it's done. Before you publish, **What publishing changes** lists which Slack workspaces will start using the new version and which are pinned to an older one and won't change. Testing usually isn't required, but two things can stop a publish: - If the workspace turned on **Require passing tests before publishing** in [AI settings](https://agentlab.cresclab.com/docs/en/studio/ai-settings.md), you need a finished test run in which every question passed. - If a question passes on the live version but fails on this draft, you're asked **Publish even though answers got worse?** Select **Keep fixing** to go back, or **Publish anyway**. After publishing, the status changes from **Not published yet** to **Live — version N**. Later edits become a new draft, and the live version keeps serving until the next publish. > [!WARNING] Some changes skip the draft > On a published shared assistant, changing the model, the name or its actions affects everyone using it right away. That's why only owners, admins and team managers can change them. ## Visibility On a shared assistant, **Visibility** decides who can see and use it: - **Everyone in the workspace**: every member. - **Specific teams**: only the teams you select. Workspace owners and admins can always see it. With no team selected, only you and the admins can see it. Teams are created on the **Team** page first. Only workspace owners and admins can change visibility. ## Share a private assistant **Share or deploy** publishes a private assistant's saved version to other people: 1. **Choose who** Under **Share with**, choose **Everyone in the workspace** or a team. 2. **Share** Select **Share**. Sharing checks a few things first: - If the workspace requires passing tests before publishing, you need a test run in which everything passed: open **Other settings** and run the tests under **Test your assistant**. - If some answers got worse than before, sharing stops. After reviewing the comparison you can select **Share anyway**. - An assistant using one of your personal skills can't be shared. Detach the skill, or ask an admin to add it to the workspace skill library. > [!WARNING] Sharing can't be undone > Once shared, it's a shared assistant: edits take effect only when someone with publishing rights publishes them, and it's no longer yours alone. The files it reads are shared with the same people. Sharing with the whole workspace takes a workspace owner or admin. To share with a team, that team's manager can do it as well. ## Version history Every save, manual edit and restore adds a version. **Version history** marks the **Current draft** and the **Live** version, and shows who built and who published each one. **Restore this version** copies an older version's instructions into a new draft, and the current draft stays in the history. On a shared assistant the restored draft still needs publishing; on a private assistant your next message uses it. ## API keys **API keys** let another system call this assistant directly, without a conversation. Under **Create an API key**, enter a **Key name** and select **Create key**. The full key is shown only once, so copy it and store it somewhere safe. After that the list shows only the key's name and when it was last used. Select **Revoke** to retire a key. Only workspace owners and admins can create or revoke keys. ## See what it costs Model spend per assistant is in **Workspace settings** › **Usage**, on the **Assistants** tab, under **Assistant model spend**. It shows estimated model cost by assistant and by reply, along with reply and model-call counts. Tool calls are counted (succeeded, failed and so on) but not given a price. Each reply is labeled by source: conversation, channel reply, automation schedule, automation trigger or assistant test. Only workspace owners and admins can open this page. ## Archive an assistant Assistants can't be deleted, only archived. An archived assistant disappears from the list and stops counting toward the limit. Its conversations and history are kept, but it can't be opened from the console again, and archiving can't be undone from the console. 1. **Open the section** Under **Advanced and integration** (**Other settings** on a private assistant), expand **Archive assistant**. 2. **Confirm** Select **Archive assistant**, then **Archive** in the confirmation. If the assistant still answers in Slack workspaces, the confirmation names them. After archiving, nobody answers there, so assign another assistant first. ## Who can do what | Action | Member | Owner or admin | | ----------------------------------------------------------------------------- | ------ | -------------- | | Create an assistant | Yes | Yes | | Change your own private assistant, including archiving it | Yes | Yes | | Edit the draft of a shared assistant you can see, run tests, restore versions | Yes | Yes | | Change the model, name or actions of a published shared assistant | No | Yes | | Publish a shared assistant, overturn a test verdict | No | Yes | | Change visibility, share a private assistant with the whole workspace | No | Yes | | Archive a shared assistant | No | Yes | | Create or revoke API keys | No | Yes | | See assistant spend | No | Yes | For an assistant whose visibility is **Specific teams**, a manager of one of those teams can also publish and archive it, and change the model, name and actions of its live version. A team manager can also share their own private assistant with that team. Every workspace has a general assistant created by Agent Lab. Only owners and admins can change its name and instructions. It can't be archived and has no model setting of its own.