Agents

An agent is the smallest unit of work in the agentic framework: one versioned prompt plus one model, and optionally the tools and skills that prompt is allowed to call. Nothing more. An agent does one job with one set of instructions.

What an agent is made of

The prompt is versioned, so you can change instructions over time and see what an agent looked like at any point. The model is the LLM the prompt runs against. Tools and skills are optional; attach them when the agent needs to read or write beyond plain text. You create and edit agents in the Developer group under Agents.

Model assignment and overrides

There are three ways a model gets chosen, from most general to most specific:

• Platform default. An agent can inherit the platform default model. When an agent inherits, changing the one platform setting changes every inheriting agent at once, instead of editing each agent by hand.
• Per-agent assignment. Pin a specific model on the agent itself. This overrides the platform default for that agent only.
• Per-call override. Pass a model at call time. This overrides both the agent's assignment and the platform default for that single run.

The pattern that saves the most maintenance: leave most agents on the platform default, pin a model only where an agent genuinely needs a different one, and reserve per-call overrides for one-off experiments. See Prompts for how versioned prompts fit in.

Agent versus assistant

Keep these distinct. An agent is a single unit: one prompt, one model. An assistant is the composite deliverable: several agents working together, usually with a dashboard, to cover an end-to-end job. When the quickstart says you are building an assistant, it means the whole thing, not a lone agent.

Agents are driven through the AI/Agents API. For exact request shapes and the current model list, open the live interactive reference from the app's user menu.