Agent

A task requires several dependent steps:
Find the incident → inspect logs → identify the service → check deployment → summarize findings.

Each result determines what to investigate next, so the system needs to choose and execute further steps.

The model generates responses or requests tool calls. An agent combines that model with tools and software that executes calls, returns results, and lets the model choose the next step toward a goal. The same model can power ordinary chat or an agent. Chat can also use tools and multiple steps; agentic behavior means the model directs the next actions based on results, within the application's permissions and limits.

Reason
  ↓
Tool
  ↓
Observe
  ↓
Reason
  ↓
Next tool