Create AI Agent in Bold Agent

Bold Agent helps teams build AI agents with clear instructions, governed knowledge (Libraries), and optional tool integrations (MCP servers). You can create agents for scenarios such as customer support, internal knowledge, and workflows.

This article shows you how to create an agent in Agent Builder, configure it, test it in the Playground, and then publish and share a version through Agent Hub or a website widget.

1.Create an agent

1. Go to Agent Builder.
2. Select Create agent.

3. In the Create Agent form, fill in:

   • Name: Enter a name for your agent.
   • Description (required): A clear summary of the agent’s purpose and the given description will be shown in the agent hub.
   • Model (required): Select the AI model.
   • Access (required): Choose who can see and build the agent.
     • Private: Visible only to you.
     • Teams: Visible to selected teams.
     • Global: Visible to anyone in your workspace/organization.

   When you select Teams, an additional Teams field appears. Select one or more teams to control who can discover and build this agent.

   How Teams access works

   • Only members of the selected teams can discover and open the agent in Agent Builder.
   • Selected team members can work on the agent collaboratively by updating instructions, changing settings, and linking resources such as Libraries and MCP servers.
   • Selected team members can test the agent in the Playground and refine the configuration before publishing.

4. Select Create.

After creation, the agent opens in Draft mode (for example, v0.1 (Draft)), where you can configure and test it before publishing.

2.Add agent instructions

Instructions define the agent’s role, boundaries, response style, and expected output format. Clear instructions are the fastest way to make responses consistent.

1. In the left pane, open Instructions.
2. Add comprehensive instructions such as:
   • Role and audience (who the agent is)
   • Scope (what it can and cannot answer)
   • Tone and formatting rules
   • Safety and escalation guidance (when to ask for clarification or hand off)
3. Select Update to save changes.

3.Link Libraries (optional)

Libraries provide governed knowledge that your agent can reference when responding to users. Linking a library helps the agent answer from approved, up-to-date content (instead of relying only on general model knowledge), which improves accuracy and reduces inconsistent or unverified responses.

Use a library when you want the agent to:

• Answer based on official sources such as product documentation, internal policies, SOPs, and curated Q&A.
• Follow your organization’s terminology, definitions, and support process.
• Avoid guessing when information is missing or ambiguous.

Common library sources typically include documents (for example, PDFs), web pages/URLs, and canonical Q&A entries.

Follow these steps to link a library to the agent:

1. Open Libraries.
2. Select Link library.
3. Choose the library (or create one first), then link it to the agent.

After linking, validate in the Playground by asking a few questions that should be answered directly from the library.

4.Link MCP servers (optional)

MCP servers let the agent call tools (for example, internal systems, workflows, or external APIs) through a secure MCP endpoint.

1. Open MCP Servers.
2. Select Link MCP server.
3. Choose the MCP server and confirm the tools you want to expose to the agent.

Use MCP when you want your agent to take actions (not just answer questions).

5.Configure agent settings

Settings let you select the model and tune runtime behavior such as error handling, generation parameters (for example, Temperature and Top P), conversation context, retrieval controls, and any model-provided built-in tools.

1. Open Settings.
2. Configure the following (availability may vary by model):

Model

• Model: Change the model used by the agent.

Error handling

• Custom error message: The message shown when the agent encounters an error.

Response behavior

• Temperature: Higher values produce more varied/creative responses; lower values are more deterministic and consistent.
• Top P: Another creativity/variance control. If you use Top P, keep Temperature moderate to avoid overly random responses.
• History message limit: How many previous messages the agent keeps in context for the conversation.
• Fallback instruction: The message shown when the AI agent cannot provide a response to a user’s request.

Built-in tools (model-dependent)

If you use a model that supports built-in web tools (for example, Gemini, Azure OpenAI), you may see options such as:

• Google Web Search: Allows the agent to use publicly available web results.
• Google URL Context: Allows the agent to fetch and use content from a provided URL.

Important: When web tools are enabled, the agent may answer using web content instead of your connected Libraries or MCP tools. Keep these options off when you need responses strictly from governed sources.

Retrieval settings

If retrieval controls are available, configure:

• Chunk size: How content is split for retrieval.
• Relevance threshold: How strictly the system filters retrieved content.

3. Select Update to save your configuration.

6.Test in the Playground

Use the Playground to validate:

• The agent follows your Instructions (tone, format, and boundaries).
• Responses are grounded in Libraries when linked.
• Tool calls work correctly when MCP servers are linked.

7.Publish the agent

When the agent behaves as expected, publish the current draft as a new version.

1. Select Publish.
2. Review the version that is ready to be released.
3. Click Publish vX.0 (for example, Publish v3.0).

When you publish:

1. The current draft (for example, v2.8) is promoted to the next major version (for example, v3.0).
2. The newly published version becomes active immediately.
3. The previously published version is automatically unpublished and retained for reference.
   Note: Only major versions (v2.0, v3.0, etc.) can be published. All changes are maintained as drafts until they are published.

8.Share the agent

After publishing the agent, you can share it and control who can access it.

1. Select Share.
2. In the Share dialog, choose where you want to make the agent available:
   - Widget: Embed the agent into a website or web application using a widget.
   • Agent Hub: Make the agent accessible to users within your organization.

Share the agent via Widget

Use this option when you want to embed the agent directly into your website or customer‑facing application using a chat widget.

1. In the Share dialog, choose Widget.
2. If the widget option is not available, create one by following the steps below.

Create a Widget

If you want to deploy the agent on your website, you can create or configure a Chat Widget and customize its appearance to match your brand.

1. In the left pane, open Widget.
2. Under Chat Widget, configure the widget UI:
   • Welcome text: The greeting displayed at the top of the widget.
   • Placeholder text: The hint text shown in the message input box.
   • Brand color: Select a brand color or enter a custom hex value.
   • Text color for brand: Choose the text color used on top of your brand color (for contrast/readability).
3. Under Launcher type, choose how the widget launcher should appear on your site.
4. Configure launcher details:
   • Launcher text: The label shown on the launcher (when applicable).
   • Upload launcher icon: Upload an icon to use as the launcher image.
   • Launcher position: Choose where the launcher appears on the page (for example, right).
5. Under Widget display theme, choose a theme (for example, light or dark).
6. Select Create to save and create the widget.

Add the widget to your website

1. Copy the provided embed script or installation snippet.
2. Add the embed script to your website’s HTML (usually before the closing tag).
3. Save and republish your site.

Once implemented, the widget will load your published agent, allowing users to interact with it directly from your site.

Share the agent through Agent hub

Use this option when you want users in your organization to discover and use the agent through Agent Hub, without embedding it externally.

1. When you select Agent Hub, configure the visibility settings:
   • Enable Shared publicly to make the agent discoverable in Agent Hub.
   • Choose who can access it:
     • Global: Accessible to all users in your organization.
     • Teams: Accessible only to members of the selected teams.
2. Select Publish agent to apply the sharing settings.
3. After publishing, the agent becomes visible in Agent Hub to the users or teams you selected.

This option is ideal for internal knowledge assistants, workflow‑automation agents, and private, organization‑wide AI tools.