Prompts

A prompt is the set of instructions that defines how an agent behaves. Prompts live in a central library so you write them once, reference them from the agents that need them, and update them in one place. You manage prompts in the Developer group.

Versioned building blocks

Every prompt is versioned. When you edit a prompt, the change is recorded as a new version, so you can see what the instructions said at any earlier point and understand how behavior shifted over time. Because an agent references a prompt rather than embedding its own copy of the text, the prompt is a shared building block, not a one-off string buried in a single agent.

Central management, no redeploy

The reason to keep prompts in the library is decoupling. An agent points at a prompt, and any code that calls that agent points at the agent, not at the instruction text. When you refine a prompt in the library, every agent that references it picks up the change. You do not touch or redeploy the consuming code to adjust how the agent behaves. This keeps instruction changes fast and reviewable, and keeps behavior separate from wiring.

Prompts and agents

An agent is a versioned prompt plus a model. The prompt supplies the instructions; the agent binds those instructions to a model and, optionally, to tools and skills. See Agents for how models are assigned and overridden.

For exact request shapes, open the live interactive AI/Agents API reference from the app's user menu.