Skip to main content

​
Steps

1

Open the Agents page

Select “Agents” in the sidebar and click “New agent.”
2

Enter the basics

  • Name (required) — e.g., Sales Analyst, up to 120 characters. A name an existing agent already uses is rejected (case-insensitive); the same check applies when renaming later
  • Description — a one-liner describing this agent’s role, up to 2,000 characters
  • Model — the Claude model used for new chats with this agent
  • System prompt (optional) — behavior, tone, and constraints, up to 50,000 characters. If left blank, the default system prompt is used
Each text field shows a live “n / max characters” counter, and the same limits apply when editing the agent later.
Enter a name and press “Auto-generate” to have the AI draft a description and system prompt from what you have entered so far.
3

Create

Press “Create” to create the agent and go to its detail page. Next, attach skills, connectors, and knowledge (Attach resources).

​
Creating from chat

You can also create an agent without leaving chat. In a chat with no agent selected, just ask — “Create an agent for sales analysis” — and the AI drafts the setup and shows an “Agent creation proposed” card.
  • Review the name, description, model, and system prompt on the card
  • Press “Create” to create the agent, or “Dismiss” to pass — nothing is created until you approve
  • Use the “Open” link on the card to jump to the new agent’s detail page and continue setup
  • The created agent appears in the sidebar’s agent list right away

​
Notes on choosing a model

Model changes apply to new chats only. Existing chats keep running on the model they were created with. To keep the response style consistent, we recommend sticking with the model chosen at creation time.
Older generations (Claude Opus 5 / Claude Sonnet 5 / Claude Haiku 4.5, and Opus 4.8 / Sonnet 4.6 before them) are not offered when picking a model, and an agent saved on one is not left there: every new chat runs on the newest generation of the same line (an agent saved on Claude Sonnet 4.6 answers with Claude Sonnet 5.5), the model setting shows the model it now runs on, and saving the form without touching that setting does not count as a change. Chats that were already open keep the generation their session started on until you switch them — see When the agent’s settings change mid-chat. There is nothing to act on; the first time the composer’s model badge names a different model than you last saw on that device, a one-time callout reads “Your AI model moved from <previous> to <model>.”

​
Writing a system prompt

Put in the system prompt anything you always want the agent to follow.
  • Role: “You are an analyst responsible for sales data analysis”
  • Tone and language: “Always respond in polite, formal Japanese”
  • Constraints: “Always cite a source for every number”
  • Procedures: routine workflows
Your prompt is layered on top of the platform’s base prompt, not a replacement for it. It wins on role, style, tone, and format, while platform behavior (tool usage, memory, file handling, confirmation before destructive actions) still applies — no need to restate any of that. The response language changes only if you say so explicitly (e.g., “always answer in English”); writing the prompt itself in a given language doesn’t change it.
For concrete procedures you want to reuse, register them as a skill rather than writing them into the system prompt. Skills can be shared across agents and can receive update suggestions from chat.

​
Setting an avatar

From the Settings section of the agent’s configuration panel, you can upload an avatar image (PNG / JPEG / GIF / WebP, up to 2 MB). It helps you identify the agent in lists and chat screens.

​
Built-in document creation skills

Every agent supports creating PowerPoint (.pptx), Word (.docx), Excel (.xlsx), and PDF documents out of the box. No extra setup is required.