OpenAI Debuts Decisions API to Monitor AI Agents

During its recent Dev Day event, OpenAI CEO Sam Altman announced a limited preview of the Decisions API, a new tool designed for software automation. The API allows developers to provide the company's Luna model with a predefined set of choices, such as image classification categories or specific agent behaviors. By narrowing the model's focus to these specific options, OpenAI aims to deliver extremely fast outputs while retaining core capabilities like image understanding, multilingual support, and safety guardrails.

The Decisions API closely mirrors Jev, a fast classifier model released earlier this month by TypeSafe AI. Founded by former OpenAI engineer Diogo Almeida, TypeSafe AI designed Jev to handle what it calls "System One" tasks—fast, intuitive decision-making—rather than the slower, more expensive "System Two" deliberate reasoning typical of large language models. Almeida noted that while making cheap and fast models is easy, the real challenge lies in maximizing intelligence per dollar, which TypeSafe achieves using synthetic data.

For AI practitioners, these decision-focused APIs offer a highly efficient way to monitor and secure autonomous agents. Currently, guarding against rogue agent behavior requires running separate, expensive frontier models to watch for bad actions at a significant compute cost. Shapor Naghibzadeh, a cybersecurity professional who leads the startup QueryStory, demonstrated this potential by building a hackathon project that uses Jev to evaluate agent actions against assigned tasks. His system blocks actions with high confidence of being bad, flags others for review, and permits the rest. While monitoring these actions with a standard frontier LLM costs $372, doing so with Jev costs just $2.94, making continuous oversight financially viable and potentially preventing major security incidents.