Skip to main content
When an Agent’s answer is wrong, either the tool returned something wrong, the Agent picked the wrong tool, or it called the right tool with the wrong arguments. The tool spans (wrench icon) in the span tree show all three.

​
Tool span detail

Click a tool span and four sections appear.
  • Tool The definition of the tool that ran: name, runtime, description, and for an inline SQL tool the exact query body.
  • Input The arguments the model chose, as key-value pairs.
  • Output The tool’s raw response. Row data renders as a table, so it is easy to scan.
  • Metadata The tool context (toolContext) at call time.
The Formatted / JSON toggle also gives you the raw JSON.

​
Debugging order

  1. Read the input. Wrong arguments usually trace back to vague parameter descriptions.
  2. Read the output. If the call succeeded but the data looks off, look at the tool implementation.
  3. Read the tool definition. If the Agent picked the wrong tool, tighten the tool description. Agents choose by description.

​
Failed tool calls

A run whose tool call failed lands in the Error status in the list, and the failure is captured as a runtime-error signal on the Signals page. See the signals overview.