self-reflection agents

A self-reflection agent produces an answer, then turns its own attention back on that answer to find flaws and revise it, possibly over several rounds. The intuition is the human habit of re-reading what you wrote: the first pass gets the ideas down, and a critical second pass catches the contradiction, the missing case, or the unjustified leap. The agent plays both author and reviewer.

Mechanically this is a generate–critique–revise loop on a single model (or a generator paired with a separate critic). The model first emits a candidate, then is re-prompted to critique it against a rubric or the task spec, and finally to rewrite using that critique; methods like Self-Refine iterate until a stopping rule fires. Crucially the critique is most reliable when grounded — by running tests, executing code, or checking a tool's output — rather than relying on the model to introspect on facts it cannot verify.

Reflection helps most when verification is cheaper than generation: spotting that code fails a unit test, or that a draft violates an explicit constraint. It helps least on questions the model simply does not know, where unanchored self-critique can talk itself out of a correct answer or into a confident wrong one. The benefit is real but uneven, and each round costs additional inference.