Skip to main content
Custom agents let you build AI teammates tailored to specific roles and workflows. Here’s how to create one.

Before you start

Think through these questions before you begin:
  • What’s this agent for? Be specific. An agent with a clear role gives better results than a general one.
  • What tools should it read? Which connected apps will this agent draw from?
  • What actions should it take, and what should it wait on? See Agent Permissions.

Create a custom agent

1

Name your agent

Choose a name your team will recognize and remember. This is how they’ll tag it in conversations: @[name].Use a clear, descriptive name. A recognizable name makes the agent feel like a teammate, not a tool.
2

Set the role

Write a short role description, one or two sentences. This is what your team sees when they look up who to tag.Example: “Handles contract reviews, compliance questions, and policy summaries for the team.”
3

Write the instructions

Instructions tell the agent how to behave: what to prioritize, how to respond, what to avoid.See Agent Instructions for guidance on writing good instructions.
4

Add artifacts

Upload documents, tables, or other reference material the agent should use as context. Brand guides, policy docs, playbooks, templates, anything that helps the agent do its job better.
5

Connect apps

Give this agent access to specific apps from 500+ integrations: CRM, project management, analytics, developer tools, communication, and more. Each agent can have its own set of connections relevant to its role.A legal agent probably needs access to your policy docs and legal folders, not your CRM.
6

Save and test

Save the agent, open a Space where it’s available, and tag it with @[name] to test it with a real question.
You can have multiple agents in the same Space. Your team tags whichever one they need with @, and each agent responds from its own instructions, artifacts, and connected apps.

Tips for a useful agent

  • Give it a specific role, not a general one. “Answer any question” is less useful than “summarize customer support threads and surface patterns.”
  • Test with real questions. Use context your team would actually ask about to see how the agent responds.
  • Iterate on instructions. The first version rarely gets everything right. Refine based on what your team actually needs from it.