Simplicity is Enough: ReAct Agents for Prompt Optimization

Prompt optimization has recently been approached with increasingly specialized search procedures and evolutionary frameworks. In this work, we ask whether such complexity is necessary to achieve competitive performance. We introduce REAPO (ReAct Agents for Prompt Optimization), a lightweight framework that casts prompt optimization as an agentic process built around a ReAct loop augmented with evaluation and reflection tools. Rather than relying on bespoke optimization machinery, REAPO iteratively refines prompts through error analysis, reflective reasoning, and validation against the optimization objective. We evaluate REAPO on several established benchmarks spanning multi-hop question answering, mathematical reasoning, fact verification, instruction following, and structured classification, as well as on real enterprise agentic tasks. Across these settings, we find that this agentic optimizer can match or outperform specialized prompt tuning methods while requiring 2–7$\times$ fewer evaluation rollouts and less auxiliary optimization logic. We also report confidence intervals over multiple independent runs and show that single-run evaluations on small benchmarks can give a misleading picture of relative performance, since LLM variance may obscure true gains. Taken together, our results suggest that effective prompt optimization need not depend on increasingly elaborate optimization schemes, and that agentic methods can provide a more efficient alternative.