OpenAI’s new API to give AI agents a watchdog
Decisions API allows a model to choose between a predefined set of options
OpenAI is testing a new API that could give AI agents a faster and cheaper way to make simple decisions, potentially helping developers keep increasingly autonomous AI systems under control.
The Decisions API, introduced by OpenAI CEO Sam Altman during the company's Dev Day event, allows a model to choose between a predefined set of options. Rather than asking a large language model to generate a full response, developers can use it for narrower decisions such as classifying an image or selecting how an AI agent should behave.
According to Altman, the API could enable the model to concentrate on a particular choice, making it work rapidly but having such abilities as image comprehension, language versatility and safety measures at its disposal.
This methodology is similar to Jev, which was recently introduced by TypeSafe AI. Jev is built to serve as a rapid classifier that assigns probability values to the range of possible options rather than producing long outputs.
As suggested by the CEO of TypeSafe, Diogo Almeida (who used to be an engineer at OpenAI), such interest from OpenAI may indicate increasing demand for this kind of solution.
Large language models can be expensive and relatively slow when used for every small decision an AI agent needs to make.
That has led developers to combine larger models with smaller, specialised systems that can handle simpler tasks. The potential benefit is lower cost and faster responses without requiring a frontier model for every step.
OpenAI's Decisions API is currently only available as a limited preview, so it is not yet clear how closely its performance will match Jev or other similar systems.
Another possible usage scenario could be watching AI agents while working on the Internet.
After some unexpected behaviour incidents with agents, OpenAI have said that they use a second model to watch their actions, but they have noted that this technique could require considerable computing power expenses.
Shapor Naghibzadeh, the head of the cybersecurity startup QueryStory, developed a demo application using Jev to examine the actions of each agent concerning its task.
It will prohibit actions that are definitely dangerous, report suspicious ones for a human to decide, and allow the rest of the actions.
Naghibzadeh's demonstration suggests why fast decision models could become useful as a security layer around AI agents. He estimated that monitoring a particular scenario with Jev would cost about $2.94, compared with roughly $372 using a frontier language model.
