Cognition · Software component
Reflection Critic
Software componentCognitionCognition & MemoryVariation point (abstract)arc:ReflectionCritic
An abstract cognition component that critiques a generated output against evaluation criteria, identifies errors or gaps, and triggers a refined generation.
Responsibility. Critiques outputs and triggers their revision.
Also known as: Critic, Self-critique
Variants
| Variant | When to choose |
|---|---|
| Dual-Agent Critic | Choose when reducing self-reinforcing bias justifies added complexity and cost; requires critic criteria aligned with downstream consumers. |
| Self-Reflection Critic | Choose when implementation simplicity matters and the generating model has adequate domain knowledge; beware it may reinforce its own misconceptions. |
| Stepwise Reasoning Verifier | Choose when errors must be caught before they propagate through the chain; substantially improves final reasoning quality compared to single-pass generation with post-hoc evaluation. |
Relationships
is configured by structural
invokes dependency
is invoked by dependency
writes dependency
- Working Memory Buffer abstract Ch1.2
is triggered by dynamic
triggers dynamic
is orchestrated by control
evaluates assurance
Design guidance
- MUST apply robust stopping criteria to avoid over-reflection loops.
- SHOULD limit refinement to about two to three iterations, beyond which gains plateau.
- SHOULD ground critique with external tools, knowledge bases, or human verification rather than self-knowledge alone.
- SHOULD NOT be used for latency-critical, high-volume budget-constrained, or objectively verifiable tasks.
Quantitative guidance
As stated by the sources; verify before use.
- Quality gains plateau after two to three reflection iterations while cost and latency increase linearly (Ch1.2).
- Reflection turns one generation into three to five or more (Ch1.2).
- High-performing SWE-Bench agents incur 10-50x more tokens per task due to iterative reasoning and reflection (Ch1.2).
- Real-time customer service cannot afford 5-10 seconds of reflection before responding (Ch1.2).
Classification
- Patterns
- Reflection (Generate-Reflect-Refine)Scheduled / post-action / quality-triggered reflection
- Quality attributes
- Functional suitability: correctness and validity (ISO/IEC 25010 | NIST AI RMF: valid)
- Risks mitigated
- Uncorrected output errors
Sources
- Ch1.2: T. Nguyen, "Core Agent Patterns," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 1.2. ISBN: 9798244538229.
- Ch5.2: T. Nguyen, "Tree-of-Thought (ToT) Fundamentals," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 5.2. ISBN: 9798244538229.