Critic and Validator: The Quality Filter

The #1 problem with agents is that they are "Overconfident." If an agent generates something incorrect, it usually doesn't know it's incorrect.

The Critic and Validator pattern solves this by introducing a "Negative" persona whose sole job is to point out mistakes.

1. The Adversarial Concept

The Generator: Creates the content (e.g., Code, Blog post).
The Critic: Finds errors, gaps, and logical flaws.
The Loop: The Generator must update the content based on the Critic's feedback until the Critic is satisfied.

This is exactly how Generative Adversarial Networks (GANs) work in classic AI, but applied to LLM reasoning.

2. Validator vs. Critic

Critic (LLM): Qualitative feedback. "This tone is too aggressive."
Validator (Code): Quantitative feedback. "The code failed the unit test." or "The JSON is missing the 'id' field."

A production agent system uses both.

3. Visualizing the Loop

graph LR
    Draft[Generator: Create Draft] --> Critique[Critic: Find Flaws]
    Critique --> Loop{Is it perfect?}
    Loop -- No --> Draft
    Loop -- Yes --> Final[Publish]

4. Why This Pattern is Essential

A. Reducing Hallucinations

Generators often hallucinate facts. Critics, when given a "Source document" (Module 10 RAG), are excellent at spotting when the Generator says something that isn't in the source.

B. Enforcing Style

If you are writing for a brand, a Critic can be programmed with the "Brand Voice Guide" and reject any text that uses forbidden words.

5. Code Example: The "Critique" Prompt

SYSTEM """
You are a Senior Editor. You are review the following text for:
1. Logical consistency.
2. Grammar errors.
3. Fulfillment of user requirements.

Provide a list of requested changes. If the text is perfect, write 'APPROVE'.
"""

USER """
[Generate Text Here]
"""

6. The "Consensus" Method

In high-stakes environments, you can use Multiple Critics.

Critic 1 looks at Security.
Critic 2 looks at Performance.
Critic 3 looks at Readability. Only when all three say "APPROVE" is the task considered done.

Key Takeaways

The Critic pattern creates a self-correcting feedback loop.
It is easier for an LLM to spot a mistake than to avoid making it in the first place.
Validators (scripts) are faster and cheaper for structural checks.
This pattern is the key to achieving 99% accuracy in agentic systems.

Module 3 Lesson 4: The Critic and Validator Pattern