Scientific workflows that use AI agents can finish successfully and still produce unreliable results. According to a new preprint, the danger is not always a dramatic failure but a quiet one: a silent numerical error, a stale input, or a defect in a shared upstream step. Such flaws can contaminate several outputs while a sibling branch of the workflow remains unaffected, making the problem hard to spot.

The paper's title, "Structure Before Intelligence," points to its proposed remedy: gating what agentic workflows release. Rather than relying on the agents' judgment alone, the authors argue for structural checkpoints that filter outputs before they are let out. This shifts the focus from making agents smarter to making their release process more disciplined.

This is a single preprint, so the claims are not yet peer-reviewed or tested in practice. Still, the framing highlights a growing concern in computational science: as workflows become more autonomous, ensuring they fail loudly rather than silently may be as important as improving their intelligence.