---
title: Agents
sidebar_position: 1
---

# Agents

:::warning
Agents are a legacy Chat feature. The Chat tab is being removed progressively — Enterprise organizations may still have access during the six-month migration window.

If you no longer have access, use [Skills](/vibe/work/skills) instead: reusable instructions, tools, and knowledge sources that you configure once and reuse across tasks.
:::

An Agent is a **specialized assistant** you configure once and reuse across conversations. Give it instructions, tools, and knowledge sources, and it handles the rest.

:::tip
When you catch yourself **repeating the same prompts**, attaching the same files, or enabling the same tools every time you start a conversation, that's a good sign you need an Agent.
:::

<SectionTab as="h2" sectionId="use-cases">Common use cases</SectionTab>

Agents work best when you need the same instructions, tone, or knowledge applied consistently:

- **Brand voice editor**: paste any draft and get it rewritten in your company's tone, with the same formatting rules every time.
- **Legal clause checker**: upload a contract and ask the Agent to flag non-standard clauses based on your internal policy [Library](/vibe/work/libraries).
- **Meeting prep assistant**: share a prospect's website or a briefing doc and get structured talking points, key figures, and open questions.
- **Technical FAQ**: attach your product documentation as a [Library](/vibe/work/libraries) and share the Agent with your team so anyone can get accurate, sourced answers.
- **Translation with context**: an Agent that translates content while respecting your glossary and brand terminology, attached as a [Library](/vibe/work/libraries).

<SectionTab as="h2" sectionId="creating-an-agent">Creating an Agent</SectionTab>

Creating an Agent is simple and accessible to all users:

1. Select the **Chat** tab in the sidebar.
2. Click `Agents` in the left sidebar.
3. Click `Create Agent` to start from scratch, or pick a **pre-built template** as a starting point.
4. Configure the Agent (see [next section](#configuration)).
5. Test it in the **preview panel** on the right side of the screen.

Agents save automatically as you configure them. There's no save button.

:::tip
Templates cover common patterns like summarizers, translators, and code assistants. Start with one and customize it rather than building from scratch.
:::

<SectionTab as="h2" sectionId="configuration">Configuration</SectionTab>

Agent configuration is split into **main settings** and more advanced options. You can always come back and edit these later.

<ExplorerTabs>
<ExplorerTab value="main-settings" label="Main settings" default>

Every Agent starts with four fields.

- **Name**: use a clear, descriptive name your team can find at a glance. For example, `Customer Support Triage` or `Sales Brief Generator`.
- **Description** (optional): a short explanation of what the Agent does, visible in the Agent picker.
- **Avatar** (optional): choose from presets or generate a custom one.
- **Instructions**: detailed guidance for the Agent's behavior. This is where you define its role, scope, and output format. Type `/` in the field to reference tools and knowledge sources inline.

</ExplorerTab>
<ExplorerTab value="guardrails-tone" label="Guardrails and tone">

These settings are optional but can **refine how the Agent communicates**.

- **Guardrails**: set boundaries the Agent should respect. For example, *"Always use metric units"* or *"Don't share pricing information outside the approved rate card."*
- **Tone and personality**: expand `Tone` to define the Agent's communication style. A customer-facing Agent might need a warm, professional tone, while an internal code review Agent can be more direct.

</ExplorerTab>
<ExplorerTab value="tools-knowledge" label="Tools and knowledge">

Agents can be enhanced with **pre-configured tools** and your own **knowledge sources**.

**Tools**

- **[Web search](/vibe/work/web-search-open-url)**: retrieve current information from the web.
- **[Code Interpreter](/vibe/chat-legacy/code-interpreter)**: run Python for calculations, data analysis, and charts.
- **[Canvas](/vibe/work/files-and-canvas)**: generate and edit documents, code, spreadsheets, or presentations.
- **[Image generation](/vibe/work/image-generation)**: create images from text descriptions.

**Knowledge**

- **[Libraries](/vibe/work/libraries)**: attach one or more Libraries to give the Agent access to your uploaded documents.
- **[Connectors](/vibe/work/connectors)**: attach Connectors to pull data from external services like Google Drive, GitHub, or Notion.

</ExplorerTab>
</ExplorerTabs>

<SectionTab as="h2" sectionId="using-agents">Using Agents in Chat</SectionTab>

<SectionTab as="h3" variant="secondary" sectionId="starting-conversation">Starting a conversation with an Agent</SectionTab>

1. Select the **Chat** tab in the sidebar.
2. Click `+` then select `Agents` or type `/` in the message box.
{/* TODO: It's still @ in the older UI */}
3. Search for or select an Agent. Its name appears at the top of the chat box.
4. Send your message. The Agent responds according to its configuration.

<SectionTab as="h3" variant="secondary" sectionId="switching-agents">Switching Agents</SectionTab>

Click the dropdown arrow next to the Agent name, or repeat the process above, and select a different Agent.

:::info
Only **one Agent can be active at a time** within a conversation.
:::

<SectionTab as="h3" variant="secondary" sectionId="returning-default-model">Returning to the default model</SectionTab>

Click the cross icon next to the Agent name to remove it. The conversation continues with the default Vibe model. You can also start a new chat, which always begins with the default model.

<SectionTab as="h2" sectionId="writing-instructions">Writing effective instructions</SectionTab>

Good instructions make the difference between a generic chatbot and a reliable team tool. Here are a few principles:

- **Define the role clearly.** Start with who the Agent is and what it does: *"You are a code reviewer. Focus on Python best practices, security issues, and performance."*
- **Specify the output format.** Tell the Agent whether you expect bullet points, JSON, prose, a table, or something else. Include a short example if the format matters.
- **Set explicit boundaries.** State what the Agent should not do: *"Don't provide legal advice"* or *"Only reference documents from the attached Library."*
- **Reference tools inline.** Type `/` in the instructions field to mention specific tools or knowledge sources. This tells the Agent when and how to use them.
- **Iterate in small steps.** Change one thing, test in the preview panel, repeat. Getting instructions right is an iterative process.

:::info
For a broader introduction to prompt writing techniques, see our [prompt engineering guide](/inference/prompting).
:::

<SectionTab as="h2" sectionId="sharing">Sharing and permissions</SectionTab>

By default, a new Agent is **private** (only you can see and use it).

<SectionTab as="h3" variant="secondary" sectionId="sharing-specific-people">Sharing with specific people</SectionTab>

1. Open the Agent's card on the `Agents` page.
2. Click the sharing status label (`Private`, `Organization`, or `x people`).
3. Search for users or groups and assign a permission level:
   - **Collaborator**: can use, edit, and delete the Agent.
   - **Viewer**: can use the Agent in their own conversations but can't edit or delete it.

Changes save automatically. Close the modal when you're done.

<SectionTab as="h3" variant="secondary" sectionId="sharing-organization">Sharing with your entire organization</SectionTab>

In the same sharing modal, toggle `Entire organization` and choose whether everyone gets Collaborator or Viewer access.

:::note
We don't send notifications when you share an Agent. Let your teammates know so they can start using it.
:::

<SectionTab as="h2" sectionId="managing-agents">Managing Agents</SectionTab>

<SectionTab as="h3" variant="secondary" sectionId="duplicating-agent">Duplicating an Agent</SectionTab>

Open the three-dots menu (`⠇`) on the Agent card and select `Duplicate my Agent`. This creates a new, independent copy you own. Changes to one won't affect the other.

:::tip
Duplicating is also the workaround for transferring an Agent to a colleague. Share it as a Collaborator, ask them to duplicate it, then delete the original if needed.
:::

<SectionTab as="h3" variant="secondary" sectionId="deleting-agent">Deleting an Agent</SectionTab>

1. Open the three-dots menu (`⠇`) on the Agent card.
2. Select `Delete` and confirm.

:::caution
Deleting an Agent is permanent and can't be undone. If the Agent was shared, it becomes unavailable to everyone.
:::

:::note
Agents are available on all plans. If you change or cancel your subscription, your Agents and their sharing settings are preserved. Some tools or knowledge sources may become unavailable on lower-tier plans.
:::

<SectionTab as="h2" sectionId="agents-and-api">Agents and the API</SectionTab>

Agents created in Chat are designed for the chat interface and can't be used programmatically. If you need API access, create Agents through <AppLink href="https://console.mistral.ai/build/agents" app="studio">Agents</AppLink> or the [REST API](https://docs.mistral.ai/api/#tag/beta.agents).

Studio Agents can be deployed to Chat and appear in a dedicated section of your Agent list.

<SectionTab as="h2" sectionId="related-features">Related features</SectionTab>

- **[Skills](/vibe/work/skills)**: the Vibe Work successor for reusable, specialized behavior.
- **[Custom instructions](/vibe/work/custom-instructions)**: set persistent behavior rules that apply to all conversations (Agents override these when active).
- **[Libraries](/vibe/work/libraries)**: build knowledge bases from uploaded documents to attach to your Agents.
- **[Connectors](/vibe/work/connectors)**: connect external data sources like Google Drive, GitHub, or Notion.
- **[Files and Canvas](/vibe/work/files-and-canvas)**: create and edit content collaboratively with AI.
- **[Code Interpreter](/vibe/chat-legacy/code-interpreter)**: run Python for data analysis and visualization.
- **[Projects](/vibe/work/projects)**: organize conversations into scoped work areas.