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.