Why does Reflexion split the agent into actor, evaluator and self-reflection roles?

## Three roles, three jobs Reflexion decomposes a self-improving loop into roles that are often collapsed in practice, and the decomposition is the interesting part. **Actor.** Produces the attempt — the reasoning, the actions, the observations, the final answer. It is optimizing for completing the task, and it holds all the assumptions that produced the attempt, correct and incorrect alike. **Evaluator.** Judges the produced trajectory and returns a signal. That signal can be programmatic (a constraint check, a comparison against an expected result) or a model asked to grade against a rubric. Its output is typically sparse: pass/fail, a score, or a named violated constraint. **Self-reflection.** Reads the trajectory together with the evaluator's verdict and writes a short natural-language explanation of the likely cause plus a concrete change. Its output is what is carried into the next attempt. ## Why not let the actor judge itself The practical argument is anchoring. An actor asked, in the same pass, "is this right?" is conditioned on the chain of reasoning that produced the answer. Everything in that chain looks locally justified — that is why it was written — so the self-check tends to confirm. Giving the evaluation its own pass changes the framing: the evaluator sees an artifact to be attacked, not a plan it just committed to. It can also be given a different, stricter prompt, a rubric the actor never saw, or a narrower brief ("check only the hard constraints") that resists the actor's persuasion. The independence is partial, not absolute. If both roles are the same model, they share priors and blind spots: a misconception about how a library behaves will be invisible to both. Role separation buys framing independence, not epistemic independence, and a strong candidate says so rather than overselling it. ## Why the reflection step is separate from evaluation This is the part people usually miss. Evaluation and diagnosis are different tasks. "Score 0.3" or "violates the Monday constraint" is a judgement, and it is exactly the signal an actor cannot act on — it says where it landed, not which decision to change. The reflection role's whole purpose is signal conversion: from sparse verdict to specific instruction, grounded in the trajectory that produced it. Keeping it separate from the evaluator also keeps the evaluator honest; an evaluator asked to both grade and suggest fixes drifts toward grading its own suggested fix favourably. ## The itinerary example An agent plans a five-day trip under hard constraints the requester supplied: total budget, no departures before 06:00, and one museum that closes on Mondays. The actor writes a plausible itinerary. A separate critic pass, holding only the constraint list, reports two violations — the budget is exceeded by 12% and day three visits the museum on a Monday. Neither statement tells the actor what to do; the budget one in particular has many possible remedies. The reflection step reads both the itinerary and the violations and writes: "swap days three and four so the museum falls on Tuesday; the 12% overage is concentrated in the two internal flights, replace one with rail". That is what the next attempt consumes. ## Practical shapes In real systems these roles are rarely three separate services. They are commonly three prompts against the same model, sometimes with a cheaper model as evaluator, sometimes with the evaluator replaced entirely by a deterministic check where one exists. What matters for the interview is that you can articulate *why* they are distinct passes rather than one instruction saying "solve this, then check your work, then fix it". You should also be able to say what the split costs: three passes instead of one, roughly triple the latency and token spend on every failing attempt, which is why teams gate reflection on the tasks where a wrong answer is expensive. ## What weak answers sound like "You add a critic so the model checks itself" misses both the reason the pass separation matters and the existence of the third role. "Separating them makes the critic objective" overstates it — same model, same blind spots. The calibrated answer names the anchoring argument, the signal-conversion argument, and the residual shared-prior limitation.