It was in the way I was debugging it.
I spent hours trying to figure out why an agent kept choosing the wrong tool.
So I did what most of us would do.
I checked the prompt.
The final response.
The retrieved context.
The model.
Nothing looked wrong.
Then I replayed the execution.
The wrong tool wasn't the bug.
The bug happened 8 steps earlier.
A slightly different retrieval changed the next decision.
That changed the next tool.
That changed the next context.
By the time I reached the final answer, I was debugging the symptom, not the cause.
That completely changed how I debug AI agents.
I no longer ask:
"Why is this answer wrong?"
I ask:
"What's the first decision that diverged?"
Because every decision after that is usually just a consequence.
I think this is one of the biggest differences between debugging traditional software and debugging AI agents.
You're not looking for the first crash.
You're looking for the first bad decision.
Curious how others approach this.
When debugging an AI agent, do you start with the final output, or do you try to identify the first point where the execution diverged?