Once I said screw it, got up in the middle of the working day and played for half an hour. My head wasn't there. It was back in the same storm it had been in all morning, which means I didn't even give them the half hour. https://t.co/I0J7YVgwnU
Pydantic shipped something I think matters more than another +3 benchmark points.
Their new prompt optimizer reads real production traces, finds recurring failures, and suggests one evidence-backed prompt change at a time. Every recommendation points to the runs that justify it.
If the root cause is the model, provider, tool, or quota, it says so.
The other piece I like is managed prompts: immutable versions, canary rollouts, labels, and instant rollbacks. Prompts start looking like production config instead of text pasted into a dashboard.
For production agents, the question is “Which failure keeps repeating, what trace proves it, and what’s the smallest change that fixes it?”
#AIAgents
Wow, this Knight painting is stunning! 🤩 So peaceful & the light is just beautiful. Love discovering artists from this era - such a lovely, timeless piece. #artdiscovery#goldenage