Docs - Create Agent
The agent editor
Every agent lives in the editor. Fill in the fields, write a Screenplay (system prompt), pick a language, and test with Playground.
1. The fields
Each agent has five essential fields. Only avatar, name, and Screenplay must have content to publish.
| Field | Limit | Required to publish? | Notes |
|---|---|---|---|
| Avatar | 1 image | Yes | The agent's visual identity. Upload a profile picture or emoji. |
| Name | 128 chars | Yes | The agent's public name. Appears in agent lists and conversations. |
| Intro | 1000 chars | No | Public description of what the agent does. Shown on the agent card before you talk to it. |
| Screenplay | 50,000 chars | Yes | The system prompt. Tells the agent how to think, what tone to use, when to refuse, what its expertise is. |
| Language | en / es / zh | Yes | Set once at creation, cannot change. Determines the UI language and response language assumptions. |
2. The Screenplay — your system prompt
The Screenplay is the heart of every agent. It is a text block (up to 50,000 characters) that instructs the agent: what is its role, what tone should it adopt, what knowledge does it have, when should it defer to the user, what are its limits. Write your Screenplay in English or as a mix of English and markdown. Examples work better than abstract rules. Include guardrails, example conversations, or domain rules. The more specific and grounded your Screenplay, the more predictable your agent's behavior.
3. Language is set once
When you create an agent, you choose a language: English, Spanish, or Chinese. This choice is locked in and cannot be changed after creation. The language setting affects the UI language shown to the user and shapes the agent's response assumptions. If you need a different language, create a new agent in that language.
4. Test as you go
The Playground pane appears to the right of the editor. Type a message and hit send to test your agent immediately — no need to save first. Watch how it responds to different prompts, refine your Screenplay, and test again. Use the Playground to catch unintended behavior, overly strict rules, or tone mismatches before you publish.