OpenAI's Agents API adds async functions and a beta per-session budget control

An OpenAI engineer on the Agents API team listed the team's latest additions:

• Async functions. An agent can keep working after it calls a tool, without waiting for the result, for example to start slow lookups early, answer independent parts of a request, and return results when the application has them.
• Budget control (beta). Limit, monitor and control an agent's budget per session, so one customer cannot use up all of an account's tokens. Access is by request to the engineer.
• 12-hour prompt cache. Prompts can stay cached for 12 hours; the engineer says customers with long-running and personal agents have found savings in it.

The engineer asked for feature requests for next week.