Agents

An agent is a saved configuration: a model, instructions, tools, and optionally knowledge files, wrapped under a name. Instead of re-pasting the same system prompt every morning, you pick the agent and start working.

Building one

Open the agent builder from the chat and define:

• Name and description: how it appears in your picker.
• Model: the default model the agent runs on.
• Instructions: the standing system prompt, for example "You are our support drafting assistant. Tone: friendly, concise, always in English."
• Tools: enable web search, code interpreter, or file search for the agent.
• Knowledge: attach files the agent can search and cite, like product docs or style guides.

Save, and the agent is available from the model picker like any model.

Good agents in practice

The pattern that works: one agent per recurring job. A "release notes writer" that knows your changelog format. A "data analyst" with code interpreter on. A "support drafter" with your help-center docs attached. Narrow instructions beat broad ones.

Sharing

Agents are private by default. You can share an agent with specific people, and published agents appear in a gallery you can browse. Shared agents carry their instructions and tools along, so a team can standardize on one well-tuned setup. For team-wide use, see Team and organization.

Agents in your code

Every agent is also callable programmatically, in the OpenAI format, so the support agent you tuned in the app can answer on your website too. See the Agents API.

Cost

An agent costs nothing by itself. Conversations with it are billed like normal chats on the agent's model, plus whatever tools it uses.