Why is replaying an AI agent run still so difficult?

I've been thinking about a frustrating problem in agent development: when a run goes wrong, reproducing the failure isn't always straightforward.

You can save the conversation, but that doesn't necessarily preserve the tool responses, external state, or exact context that shaped each decision. Replay the run later, and the agent may take a completely different path.

That makes debugging surprisingly different from debugging a conventional program.

I'm curious how other builders handle this today.

What's the hardest part of reproducing a failed agent run in your stack? Is it capturing the full context, mocking external tools, restoring state, or dealing with nondeterministic model outputs?

I'm particularly interested in approaches people have actually tried and what still doesn't work well.