Agents

Agents build on Responses. Save the model, instructions, and tools once, then run that configuration with new input. Each run preserves its input and the upstream Responses result so you can inspect what happened.

​

Create and run an agent

Follow Create and run agents to save a definition, execute it, inspect run history, and manage versions. Use Responses directly when your application assembles each request. Use a saved agent when multiple runs should share a managed configuration.

​

Add capabilities

• Memory stores retain information across interactions.
• Chat connectors connect agents to messaging channels.
• Schedules run agents on a recurring schedule.
• Subagents delegate work during a response.

​

Explore examples

Build an agent with MCP tools, a research workflow with subagents, or a multichannel assistant. Use evals to check the behavior your application depends on.