LLMs don't always produce valid structured output matching the expected schema. Single-attempt validation leads to task failures even when retry would succeed.

The issues compound in multi-step workflows:

• Schema violations: LLM generates JSON that doesn't match the expected Zod/JSON Schema
• One-and-done failure: Single failed attempt terminates the entire workflow
• No learning from mistakes: Each step repeats the same errors independently
• Wasted tokens: Failed responses still consume context and cost money
• Fragile workflows: Flaky LLM outputs make agents unreliable

Implement multi-step retry with detailed error feedback and cross-step error accumulation. The agent learns from its validation failures across the entire workflow.

Pros:

• Higher success rate: 3-attempt retry significantly improves structured output reliability
• Cross-step learning: Agent avoids repeating mistakes across workflow
• Detailed error feedback: Zod errors guide LLM to specific fixes
• Better debugging: Error history provides diagnostic information
• Configurable balance: Can tune attempt count vs. cost/latency

Cons:

• Increased latency: Multiple LLM calls add delay when retries occur
• Higher cost: Failed attempts still consume tokens
• Context bloat: Error history consumes tokens if not limited
• Not guaranteed: Some LLMs struggle to correct from errors
• Complexity: Additional retry logic and error management

Mitigation strategies:

• Limit cross-step error window (last 3 errors) to control token usage
• Use caching to skip retries for repeated workflows
• Set per-step timeout to prevent runaway retries
• Log failures to improve prompts over time
• Consider using models with better structured output adherence
• Add exponential backoff with jitter for production deployments

• HyperAgent GitHub Repository - Original implementation (see src/agent/tools/agent.ts lines 424-509)
• Zod Validation Documentation - Schema validation library
• Self-Refine: Improving Reasoning via Iterative Feedback - ICLR 2024, Shinn et al.
• Related patterns: Structured Output Specification, Action Caching & Replay