OpenAI's Responses API keeps agents working while your tools run
Async tool calls, WebSocket mid-turn steering, and adjustable reasoning effort let you build agents that don't stall — and never lose their prompt cache.
Copy markdown
Async tool calls, no more stalling
The model keeps reasoning while your function and custom tools run, and each result streams back as it finishes. A slow build, database query, or web fetch no longer freezes the whole turn, so a multi-tool agent stays productive instead of blocking on its slowest call.
Steer a response mid-flight
Over WebSockets you can push new instructions into a response that's already generating, and the model folds in the correction without restarting. That's the missing piece for human-in-the-loop agents that need to change course mid-task instead of waiting for a bad answer to finish.
Change reasoning effort without losing your cache
Dial effort up for a hard step and back down for routine follow-ups, all inside one conversation, while the cached prompt prefix stays intact. You stop paying to rebuild context every time you shift the model's gear.
What to wire up this weekend
A coding or ops agent that fires off long tests and migrations and keeps planning while they run; a chat agent a user can redirect mid-answer. All three controls live on the Responses API (v1/responses) now.