Cursor Composer 2.5 prioritizes pleasant collaboration
Anysphere developer Tibor Tee announced that Composer 2.5, the agentic model powering the Cursor code editor, has been trained to be more pleasant to collaborate with. While emphasizing that pleasantness is crucial for real-world usefulness, he noted the lack of existing benchmarks for this quality and questioned what a pleasantness benchmark for AI agents would look like.
Making AI agents pleasant and collaborative rather than just maximizing raw intelligence is the next frontier for developer adoption, as human-AI friction is often more about communication and style than coding capability.
- –Pure intelligence benchmarks like SWE-bench don't capture the daily user experience of pair-programming.
- –Tuning for pleasantness likely involves optimizing the model's communication style, response length, and receptiveness to feedback.
- –Establishing a pleasantness benchmark is challenging due to the subjective nature of human-AI collaboration dynamics.
DISCOVERED
1h ago
2026-06-22
PUBLISHED
1h ago
2026-06-22
RELEVANCE
AUTHOR
tibor_tee