How Agents Fail: A Taxonomy

When an agent run ends normally, your code can make three promises:

• Every tool call was performed by your code, using a tool that you offered.
• The run stopped within the step limit, which is 25 steps for Desk.
• Any rule written inside a tool ran on every call to that tool.

The same run makes no promise about the following seven things:

• That the agent made progress during those steps.
• That what the agent reported really happened.
• That it chose the right tool and the right arguments.
• That it was still working on what the customer asked.