Self-Reflection, Critique, and Error Correction
Agents That Evaluate Themselves
Reflection as Control
Reflection introduces a second pass over an agent's own outputs, plans, or intermediate beliefs. The agent may critique its reasoning, detect contradictions, or compare alternative action sequences before proceeding.
Error Taxonomy
Useful categories include factual error, reasoning error, planning error, tool error, and objective mis-specification. Corrective mechanisms should be matched to the error type rather than applying generic re-prompting everywhere.
Best Practice
Treat reflection as a targeted diagnostic step, not an automatic replacement for rigorous evaluation.
Reflection Mechanisms
Which is the most appropriate use of reflection?
Reflection is most effective as a targeted revision mechanism.
Correct answer: Identify and correct specific weaknesses after a draft is produced
Why can self-critique improve agent performance on complex tasks?
A reflective pass helps catch local and global weaknesses in the initial trajectory.
Correct answer: It can expose inconsistencies or missing steps that were not obvious in the first pass.
Limits of Reflection
Reflection is not free: it consumes time, tokens, and may reinforce errors if the model critiques itself using the same flawed assumptions. Stronger systems therefore combine self-critique with external verification, test cases, or human oversight.
What is a common failure mode of reflection-only agents?
Self-critique can be circular if it is not anchored to external evidence or tests.
Correct answer: They may amplify their own mistaken assumptions