Step-by-step guide to creating an agent.
Open the Agent Creator
- Go to Agents
- Click New Agent
Name Your Agent
Give it a clear name:
image-creatorcontent-assistantdata-analyzer
Names should be lowercase with hyphens.
Choose an LLM
The AI model that powers reasoning:
| Model | Best for |
|---|---|
| GPT-4 | Complex reasoning, most capable |
| Claude | Long documents, detailed responses |
| GPT-3.5 | Fast, cost-effective |
You can change this later.
Write a System Prompt
Add Tools
Go to the Tools tab to add:
| Tool Type | What it does |
|---|---|
| App | Call apps from the grid |
| Agent | Delegate to other agents |
| Connector | MCP server tools (Linear, Slack, etc.) |
| Call | Make HTTP API calls |
| Hook | Call external webhooks |
See Adding Tools for details.
If your agent uses call tools with dynamic URLs (for example {{context.project_id}}), declare context fields so each chat can target the right resource.
Permissions (optional)
Open the permissions tab in the agent editor (shield icon in the sidebar) to set what new chats with this agent may do without asking.
| Setting | What it does |
|---|---|
| Allow every tool | Each new chat starts with every tool call allowed, remote commands included, so unattended runs (webhooks, schedules, triggers) are not blocked on approvals. Users can turn it off per chat in chat settings → Permissions. Workspace and organization policy still applies. |
These defaults are saved on the agent version (permissions.allow_all_tools). They are copied into a chat once when the conversation is created; changing the template or redeploying does not alter chats that already exist.
→ Agent template defaults · REST version.permissions
Save
Click Save. Your agent is ready.
Each save creates a new version, so you can always go back.
Test It
- Click Chat with Agent
- Try a few messages
- Watch how it uses tools
- Refine the system prompt based on results
Next
→ Writing System Prompts
→ Publishing and embedding — share the agent on external sites