How do you handle tool execution errors in agentic loops?

Short answer: Tool execution errors in agentic loops are handled by capturing runtime exceptions (timeouts, invalid JSON arguments, 4xx/5xx HTTP codes), formatting them into structured role: "tool" error messages, and appending them to the conversation context so the LLM can self-correct its next tool invocation.

Detail

If an external tool call fails (e.g. invalid SQL syntax or broken API key), crashing the python process breaks the agent loop.

[Tool Invocation Failed] ──► Catch Exception ──► Format Error String ──► Append to Agent Context
                                                                                 │
[LLM Evaluates Error]   ◄── Re-evaluate Parameters ◄── "SQL syntax error at line 1" ┘

Self-Correction Strategy

1. Graceful Exception Capture: Wrap tool execution in try-except blocks.
2. Descriptive Error Payloads: Return explicit error details (e.g. "Tool Error: Missing required parameter 'database_name'") rather than vague generic failures ("Error 500").
3. Retry Counter: Limit consecutive tool failure retries (e.g. max 3 retries per tool) to prevent infinite loops.

Example

Python agentic exception handler pattern:

def safe_tool_executor(tool_fn, tool_args: dict) -> str:
    try:
        result = tool_fn(**tool_args)
        return f"Tool Success: {result}"
    except TypeError as e:
        return f"Tool Parameter Error: Invalid arguments provided. Details: {e}. Please correct parameter keys."
    except Exception as e:
        return f"Tool Runtime Failure: {e}. Consider trying an alternative tool or approach."

Interview tips

• Emphasize that descriptive error messages allow models to learn from runtime mistakes within the same conversation loop.
• Discuss setting max step bounds to terminate failing agent tasks cleanly.

⬅ Back to AI Agents and MCP · All topics