Exception Handling & Recovery

Exception Handling ensures that an agentic system degrades gracefully rather than crashing. In the nondeterministic world of LLMs, failures are common: models hallucinate, APIs time out, and outputs are malformed. This pattern wraps critical operations in "try/catch" blocks that trigger recovery agents or fallback strategies.

When to Use

• Production Systems: Essential for any user-facing application.
• Unreliable Tools: When using 3rd-party APIs that might be down or rate-limited.
• Structured Output: When the model occasionally fails to output valid JSON.
• Safety: When a tool might return dangerous or unexpected data.

Use Cases

• API Fallback: "Primary model API failed? Switch to backup model API." or "Tool A failed? Try Tool B."
• Refusal Handling: If the model refuses to answer (due to safety filters), catch the refusal and rephrase the prompt or explain why it can't answer.
• Validation Repair: If JSON validation fails, pass the error back to the model to fix the syntax.

Implementation Pattern