Are models much better than what they seem to with current agents?

A developer observed that when using Cursor, models like Opus and Sonnet 4.6 seemed highly capable for coding tasks, with fewer "caveats" compared to their current experience with CC. This led to a question about whether models are inherently better than they appear, with their performance being either limited or enhanced by the agents running them, prompting a discussion on others' experiences.