Test an agent in the playground
Chat with it as your app would, before writing code.
Agent → Playground runs the real agent, with its instructions, tools and policy, and charges it like an API run. After each reply you'll see the run's status, latency, cost and steps.