Agent Studio

Agent Studio is a visual, drag-and-drop builder for creating custom AI agents. Build agents using a node-based canvas where you connect LLM nodes, MCP tool integrations, knowledge base lookups, web search, and conditional logic. Each agent moves through a draft/stage/production lifecycle so you can test safely before deploying to live conversations. Agent Studio builds the agents that Conversation AI and Voice AI deploy to channels.

What Agent Studio Does

The canvas provides a visual flow for how an AI agent thinks and acts. Instead of writing prompts in a text box and hoping for the best, you design the agent’s decision tree visually. Each node represents a capability: process language with an LLM, look up data from a knowledge base, call an external service via MCP, search the web for real-time information, or branch based on conditions.

This gives you precise control over what the agent can do and when. A customer service agent might start with an LLM node to understand the query, branch to a knowledge base lookup for product questions, or branch to an MCP call for order status checks. Each path is visible on the canvas, making it easy to understand, debug, and improve.

Node Types

LLM Nodes. The core processing unit. Configure the model, system prompt, temperature, and output format. LLM nodes handle natural language understanding, response generation, summarization, and classification. Chain multiple LLM nodes for complex reasoning that requires multiple processing steps.

MCP Tool Nodes. Connect to external services through the Model Context Protocol. MCP nodes can look up CRM data, create records, send messages, update pipeline stages, check inventory, or call any API exposed through an MCP server. This is how agents take action beyond conversation.

Knowledge Base Nodes. Query uploaded documents, FAQs, product catalogs, or any structured data you’ve trained the agent on. The node retrieves relevant context and passes it to the next node in the chain. Use this for product-specific questions, policy lookups, or any information the agent needs to reference.

Web Search Nodes. Enable the agent to search the web for real-time information. Useful for questions about current events, competitor pricing, or anything not covered in the knowledge base. Results feed into the next node for processing.

Conditional Nodes. Branch the flow based on variables, classifications, or outputs from previous nodes. Route angry customers to human handoff, route product questions to the knowledge base path, route booking requests to the calendar integration.

Draft/Stage/Production Lifecycle

Every agent starts in Draft mode. Build and modify freely without affecting anything live. When ready for testing, promote to Stage. Stage agents can be tested with internal users or a limited audience. After validation, promote to Production for full deployment.

This lifecycle prevents untested agents from reaching customers. If you need to modify a production agent, make changes in draft, test in stage, then promote. The production agent continues running unchanged until you explicitly push the update.

Relationship to Other AI Features

Agent Studio builds the agents. Conversation AI deploys them to text-based channels (SMS, web chat, Facebook Messenger, Instagram DM). Voice AI deploys them to phone calls. Workflow AI triggers them from workflow actions. The agent’s logic lives in Agent Studio; the deployment channel is configured separately.

This separation means you build once and deploy anywhere. A single agent can handle web chat conversations, phone calls, and workflow-triggered interactions using the same logic, knowledge base, and MCP connections.

Pro Tips

Start simple. Build an agent with one LLM node and one knowledge base node before adding MCP integrations and complex branching. Get the core conversation right first.
Use the stage environment with real test scenarios. Have team members interact with the agent as if they were customers. Note where it fails or gives unexpected responses.
Name your nodes descriptively. “Check Order Status” is more useful than “MCP Node 3” when debugging a 20-node flow six months later.
Version your agents by duplicating before major changes. This gives you a rollback path if the new version underperforms.

Common Questions

Do I need Agent Studio to use Conversation AI or Voice AI?

No. Conversation AI and Voice AI work with simpler prompt-based configurations. Agent Studio is for building more sophisticated agents with multi-step logic, external integrations, and complex decision trees.

What is MCP integration?

MCP (Model Context Protocol) allows agents to call external tools and services. Through MCP nodes, an agent can look up CRM records, check calendar availability, create tasks, update pipelines, or call any API. This is how agents move from conversation to action.

Can I revert a production agent to a previous version?

Not directly. Duplicate agents before making major changes to maintain previous versions. If the new version has issues, you can promote the duplicate back to production.

How many agents can I create?

Each sub-account supports up to 100 agents. Most businesses need 2-5 agents covering different use cases (sales, support, scheduling, after-hours).