I recently played with Microsoft Agents which are (from what I understand) powered by OpenAI.
Though, it seems that a custom GPT understands instructions much better than a Microsoft Agent seems to understand its instructions.
Has anyone else made similar observations?