DinoDS spotlights assistant routing gap

The strongest point here is architectural: if you collapse every request into generic tool use, you get exactly the kind of “nice English, wrong behavior” failure the post describes.

• –The post makes a credible case that intent routing should be separated from execution, especially for calendar, messaging, maps, and file workflows
• –A simple prompt fix is unlikely to solve this reliably; the failure mode looks more like missing supervision data for action boundaries than missing reasoning
• –The split between connector intent, connector action mapping, deeplink intent, and deeplink action mapping is the useful idea here
• –In practice, teams will probably need a layered system: rules for obvious cases, classifier/routing models for ambiguous cases, and post-training examples for edge cases
• –DinoDS is positioning itself as training data for that middle layer, which is a sensible niche if they can prove lower routing error rates