Three paths. Different control.
Choose where the agent loop runs and which infrastructure your application owns.
Where do you want control?
Agents API
Durable sessions, orchestration, and optional hosted compute.
| Responsibility | Agents API | Agents SDK | Responses API |
|---|---|---|---|
| Agent loop | OpenAI-managed harness | SDK runner in your runtime | Your application |
| State | Durable API sessions | Your session/storage integration | Your state or API conversations |
| Custom functions | Your application handler | Your application tools | Your application handler |
| Execution environment | Hosted, self-hosted, or none | Your runtime and integrations | Your infrastructure |
| Deployment | Integrate the managed service | Deploy your agent application | Deploy your orchestration |
| Best fit | Managed, continuing agent work | Application runtime control | Custom model-and-tool workflows |
These are different runtime choices. A session, sandbox, SDK runner, and model response are not interchangeable resources.
Official runtime comparison ↗