What’s the hardest production issue you’ve faced with AI agents?

Curious what kinds of failures people are seeing once agents move beyond demos and into real production workflows.

Some patterns we kept running into while building AgentPulse:

• context getting dropped between agents

• workflows degrading without throwing errors

• debugging failures across multiple tools/agents

• no clear visibility into where cost spikes happen

Would love to hear:
What’s been the hardest issue to debug in your agent workflows so far?