Are structured outputs in agents always good?
This paper suggests that you might have to take a closer look.
Your product's structured output surface is measurably more homogeneous than the chat surface you benchmarked it on.
Teams treat structured output as a formatting choice with no effect on content. This new work from Cornell tests that assumption across 44 models and finds the format itself reshapes the answers.
On an open-ended pick-a-word prompt, simply requesting JSON pushes the modal answer from 41% to 64% of responses and cuts distinct answers from 52 to 36.
The collapse shows up for JSON and XML, the formats models were post-trained to emit, and disappears for YAML and CSV. Enforcing a schema at the decoder compresses nothing further, which points the narrowing back to how the model answers the register itself.
Why does it matter?
Any pipeline leaning on JSON mode for variety is sampling from a smaller pool than it thinks.
Paper. https://arxiv.org/abs/2607.18476
Learn to build effective AI agents in our academy. https://academy.dair.ai/