Agent Builder

What the Agent Builder controls

The Agent Builder is where admins create and maintain agents. The current dashboard splits configuration into tabs for capabilities, tools, knowledge bases, threads, usage, and settings.

A typical setup starts with creating an agent and selecting a model, then adding a system prompt, tools, knowledge bases, and chat features. After that, enable the agent and grant access through groups.

Create an agent and choose a model

Open the dashboard's Agents page and click Create Agent. Creation requires a name and a model.

AgentBox supports two model sources:

• Built-in models: provided by the platform. Calls consume workspace credits based on usage.
• Custom models (BYOK): add the provider, model name, API key, and optional base URL under Settings > Custom Models, then select that model for the agent. All workspaces can use custom models. Custom model calls do not consume AgentBox credits.

An agent can use one built-in model or one custom model at a time.

System prompt and model settings

After creating the agent, use the Capabilities tab to configure the system prompt. The system prompt defines the agent's role, boundaries, behavior, and response style.

The same tab lets you switch the selected model and configure currently supported chat features.

Tool switches

Use the Tools tab to enable built-in tools per agent. Current tools include:

• Web Search: lets the LLM search for current information when needed.
• Image Gen: generates images from text descriptions.
• SOE: evaluates pronunciation from a reference text and a user audio file.

SOE supports a language setting. Built-in tools consume fixed credits per invocation. If the workspace has no remaining credits, built-in tools are not injected into that chat request.

Knowledge base association

Create a knowledge base and upload documents from the dashboard's Knowledge Bases area, then return to the agent's Knowledge Bases tab and select the knowledge bases to attach.

After association, AgentBox automatically injects retrieval, document reading, and document listing capabilities into that agent's conversations. Users do not need to turn on a separate LLM tool switch.

Voice input and voice playback

The Capabilities tab includes chat feature settings:

• ASR voice input: when enabled, sub-accounts can send voice input in chat. The current language options are Chinese and English.
• TTS voice playback: when enabled, replies can be synthesized as audio. The current settings include speed and auto-play for new messages.

ASR and TTS are chat features, not LLM-callable tools. Their calls are recorded separately for usage and credits consumption.

Publish and hand off to sub-accounts

Enable the agent from the Settings tab. A model must be configured before the agent can be enabled.

Whether a sub-account can see the agent depends on group authorization. Admins grant the agent to one or more groups and add sub-accounts to those groups. After signing in to the workspace portal, sub-accounts only see enabled agents they are authorized to use.

Admins can also copy the chat link from the agent overview or open the chat page directly to test an enabled agent.