Agent Builder

The Agent Builder is a built-in agent that designs and configures other agents through conversation. Tell it what you need, for example “an agent that answers billing questions from our refund policy and can look up an order”, and it drafts the system prompt, creates the agent, and attaches what it needs.

Every workspace has one. It is available to workspace admins.

Start a conversation

Open Agents and select Create with AI. The workspace web chat opens with the Agent Builder selected.

Describe the goal, the audience, the channels, and the systems the agent should reach. The builder asks follow-up questions when something is unclear, and tells you what it created.

What it can do

Area	Actions
Agents	List agents, show an agent’s full configuration, create an agent with web chat enabled, and update its name, description, system prompt, harness, web chat model and initial memory.
Resources	List the workspace’s tools, predefined tools, knowledge bases, MCP servers, skills and agents, and attach or detach them, including sub-agents.
Create resources	Create custom HTTP tools, empty knowledge bases and MCP server configurations.
Evaluations	Create an evaluation dataset with its cases, run it against an agent, and read each failure with the scorer’s reason.

It cannot upload documents. When it creates a knowledge base, upload the files yourself under Knowledge Base before the agent can search them.

Improve an agent with evaluations

The builder can close the loop between a prompt change and its effect in one conversation:

1. It writes cases for the behaviour you want to protect and runs them to see where the agent stands.
2. It reads which cases failed, why, and what the agent answered.
3. It writes a revised system prompt and runs the same cases against it, without saving it on the agent.
4. It compares the two results, for example 50% → 90%, and applies the new prompt only if you agree and it scored better.

Small runs finish inside the conversation; larger ones are queued like any other evaluation, and the builder reports back when you ask. It never evaluates itself.

Caution

The builder changes real agents, and its changes go live on every channel as soon as it makes them, even on an agent with versioning turned on. It confirms with you before creating or changing anything; read what it proposes before you agree. For agents already in production, prefer making changes yourself in a draft.