The Decisions API Is the Boring Win That Makes Agents Actually Work

Microsoft put the Azure OpenAI Decisions API into general availability, and almost nobody noticed. It returns choices, probabilities, and confidence scores instead of generated text. That is the most useful thing to happen to agents in months.

Agents do not need a paragraph. They need a branch to take. When a model returns 'option B, 82 percent confidence' instead of three sentences of hedging, the agent can act instead of guessing what the model meant.

The trend toward deterministic outputs, scores over sentences, is what takes agents out of the demo and into the payroll. A system that returns text needs another model to parse it. A system that returns a number just runs.

I would take ten boring API wins like this over one flashy model release. Infrastructure is what ships. The prose is what gets screenshotted.