AI agents

Create AI assistants that can answer questions, use tools, search knowledge, and help run business processes.

AI agents are assistants you configure for a project. A good agent has a clear job, an AI connection and model, instructions, and only the tools or knowledge it needs to do useful work.

Open Agents under Build.

Agents live under Build. Use New Agent to create an assistant inside a project.

In this guide

Topic	Start here
Create the agent, choose AI settings, and write instructions	Create and configure
Understand every Configure option	Configuration reference
Give the agent access to tools, knowledge, skills, and MCP integrations	Tools, knowledge, and skills
What the agent builder vs workflow builder assistants can do	Builder tools
Chat with the agent, review conversations, and share it safely	Chat and share

Setup order

Follow this order for a new agent:

Step	Why it matters
1. Create the agent in a project	The project owns the agent and its related workflows, knowledge, and permissions
2. Set AI connection and model	The agent cannot answer properly until it knows which AI provider and model to use
3. Write profile and instructions	The agent needs a job description, boundaries, and tone
4. Add tools and knowledge	Only add access the agent needs for the process it supports
5. Chat, test, and share	Test with realistic prompts before teammates rely on it

What you configure

Area	What it controls	Where
Profile & instructions	Agent name, description, avatar, and behaviour rules	Configure
AI settings	AI connection and model	Configure
Dynamic discovery	Whether the agent can adapt its setup while building	Configure
Objects	Business object types and records the agent can access	Configure
Tools	Built-in actions and external MCP integrations	Configure
Knowledge	Notebooks the agent can search	Configure
Skills	Reusable capability packs when enabled for your workspace	Configure
Sharing	Public or shared access settings	Sharing tab

Editors see the main chat tab as Build because they can both chat and shape the agent. View-only users may see it as Chat.

Good agent design

Citizen developers usually get better results by starting narrow:

One agent should own one business job, such as “triage supplier emails” or “answer policy questions”.
Give instructions that say what to do, what not to do, and when to ask a human.
Add knowledge only when the agent should answer from documents.
Add tools only when the agent needs to take action.
Test with examples before sharing with a wider group.