Skip to main content

Creating Agents

Use the New Agent Wizard to define an agent's identity, capabilities, and guardrails. The 14-step form walks you through everything from naming the agent to selecting its model and integrations.

Key Concepts

The New Agent Wizard

The wizard is a 14-step guided form that collects everything needed to create a fully configured agent. Each step validates its inputs before allowing you to proceed. You can navigate back to earlier steps to revise your choices at any time before the final submission.

Agent status after creation

All newly created agents start in draftstatus. Draft agents cannot receive work items or be dispatched. They must be activated — either directly (if the Agent Review is not enabled) or after passing agent review.

Step-by-Step Guide
1

Name and identity

Enter a namefor the agent and an optional description. The name should be descriptive of the agent's function — for example, "Support Ticket Triager" or "Invoice Processor."

Step 1 of the New Agent Wizard showing name and description fields
Give your agent a clear, descriptive name that reflects its purpose.
2

Role

Define the agent's role within your organization. This is a free-text field that describes what the agent does. The role is displayed on the agent card and used for organizational filtering.

3

Expertise

Specify the agent's areas of expertise. These are tags or keywords that describe the domains the agent is knowledgeable in, such as "customer support," "financial analysis," or "code review."

4

Model selection

Choose which AI model powers the agent. Available models depend on your configured provider keys (BYOK). Common choices include Claude, GPT, and Gemini variants.

Step 4 of the New Agent Wizard showing model selection dropdown with available models
Select the model that best fits your agent's task. Model availability depends on your configured provider keys.
Tip
For complex reasoning tasks, use a larger model like Claude Opus or GPT-4o. For high-throughput, lower-cost tasks, consider Claude Haiku or GPT-4o-mini.
5

System prompt

Write the system promptthat defines the agent's behavior, personality, and constraints. This is the foundational instruction set the model receives with every interaction.

Note
The system prompt is part of the agent's version hash. Changing it after review governance triggers drift detection and requires re-governance.
6

Integrations

Select which MCP servers and skillsthe agent can access. This defines the agent's tool surface — what external services and capabilities it can use.

Step 6 of the New Agent Wizard showing available MCP servers and skills with toggle switches
Toggle on the MCP servers and skills your agent needs. Each integration is individually controlled.

Available integrations include connected MCP servers (Slack, GitHub, Google Drive, Notion, etc.) and installed skills. Only grant the integrations your agent actually needs — the principle of least privilege applies here.

7

Environment

Choose the deployment environment: non-prod or production. This affects whether review governance is required (depending on your organization's council settings) and how the agent is labeled in the dashboard.

8

Review and create

Review all your choices on the summary screen. Verify the agent's name, role, model, system prompt, integrations, and environment. Click Create Agent to finalize.

Step 8 of the New Agent Wizard showing a review summary of all configuration choices with a Create Agent button
Review everything before creating. The agent will start in draft status.