Adaline just launched a self-improvement layer for AI agents that turns messy production traces into fresh evals， synthetic edge cases， and better agent candidates for humans to approve.

I expected it to be a regular trace viewer， but it is reading my production traffic and building evals I would never have considered.

It reads production traffic and user feedback， then clusters the mess into recognizable agent behaviours rather than asking a human to manually inspect every strange conversation.