🤖

AI Workflow Patterns Overview

Aug 11, 2025

Overview

This lecture explains two powerful AI workflow design patterns—Evaluator & Optimizer and Human-in-the-Loop—detailing their structure, use cases, and differences for improving automation reliability and accuracy.

Evaluator & Optimizer Pattern

  • One AI model (LLM) generates output, while another AI evaluates and provides feedback, repeating until the output meets criteria.
  • The workflow loops: input → generator LLM → evaluator LLM → feedback (approve or revise) → generator LLM, until approval.
  • This pattern works best when tasks have clear evaluation criteria and benefit from iterative refinement.
  • Example: In an email auto-responder, the support agent LLM generates replies, and the evaluator LLM checks for specific requirements (e.g., proper sign-off, tone).
  • Feedback is automatically structured and provided back to improve outputs each iteration.
  • Specialized evaluator agents focus on validation, improving reliability when generator prompts are complex or lengthy.

Human-in-the-Loop Pattern

  • Similar structure to Evaluator & Optimizer, but a human evaluates and provides feedback instead of an AI.
  • Used for critical decisions or tasks driven by user preference where human judgment is more reliable than current AI.
  • Example: Research report generator proposes sections, user reviews and requests changes, and the workflow loops until user approval.
  • This allows for granular control and ensures outputs meet exact human expectations.
  • Reduces unnecessary workflow cost by clarifying requirements before heavy automated processing starts.

Comparing the Patterns

  • Both patterns involve iterative feedback to refine outputs, differing only in whether AI or a human is the evaluator.
  • Use Evaluator & Optimizer for efficiency and clear criteria; use Human-in-the-Loop for reliability and subjective or critical decisions.
  • Efficiency (AI) vs. reliability/flexibility (human) often determines choice.

Key Terms & Definitions

  • LLM (Large Language Model) — An AI that understands and generates human-like text.
  • Evaluator & Optimizer Pattern — Workflow where AI generates and another AI evaluates outputs in a loop.
  • Human-in-the-Loop — Workflow where human feedback guides and approves AI outputs before finalization.
  • Feedback Loop — The process of revising outputs based on evaluator (AI or human) responses until approval.

Action Items / Next Steps

  • Practice designing an automation workflow using either the Evaluator & Optimizer or Human-in-the-Loop pattern.
  • Review workflow examples and test templates if available.
  • Identify scenarios in your projects where each pattern is best applied.