agentic-eval

Self-critique improves LLM agent outputs by implementing iterative refinement loops where the AI evaluates its content against specific criteria, identifies shortcomings, and regenerates the response to ensure quality standards are met.

An evaluator-optimizer pipeline is a system where an AI generates content, an evaluator measures response quality using rubric-based scoring or LLM-as-judge, and the optimizer refines the output based on that evaluation.

You can use iterative refinement to pass test-driven code checks by generating code, running tests, evaluating the results, and using an agentic feedback loop to critique and refine the code until it adheres to best practices and passes.

Yes, you can apply rubric-based evaluation to quality-critical generation tasks by defining structured scoring criteria, enabling the system to measure response quality and iteratively correct outputs that fail to meet the specified rubric.

You should use LLM-as-judge for measuring response quality when you need a scalable, automated method to critique complex AI outputs, applying structured evaluation strategies within an evaluator-optimizer pipeline for continuous improvement.

A limitation of self-reflection in AI code refinement is that the evaluator may share the same blind spots as the generator, requiring robust outcome-based checks and well-defined rubrics to effectively catch all functional and best practice failures.