Decision API batches and higher limits for agents

Specialists and scripts often need the same judgment over a long list: which of 20,000 leads fit a target market, which tickets in a backlog are urgent, which rows in a spreadsheet belong to which category. The decision service now takes a whole batch of items in one call and answers each one, and every built-in service accepts far more requests than before, so an agent working through a large list keeps going instead of being told to slow down.

How it works

• Batches of up to 1,000. Send states instead of state with the same questions, and each item comes back in the order you sent it. One request replaces a thousand.
• One bad item never sinks the batch. An item that fails carries its own error code, the rest still answer, and you resend only the items worth retrying.
• Billed for what was answered. A batch costs the same as asking each item on its own, and items that failed cost nothing.
• Limits sized for agents. Search, scraping, chat completion, decisions, email, and the other built-in services now allow many times more requests per minute, for workflows, apps, specialists, and calls made with a workspace API key.
• Specialists batch by default. The decisions skill now teaches specialists to send large lists in batches, a few in parallel, saving progress as they go so a rerun skips finished work.