Ceiling Effects and Convergence: Null Results for Instruction Repetition in LLM-Agent Pipelines

Abstract

Context: Prompt repetition, the verbatim duplication of an input transforming <QUERY> into <QUERY><QUERY>, has been shown to improve accuracy for non-reasoning large language models on retrieval and multiple-choice benchmarks [Leviathan et al., 2025]. Objective: We ask whether applying this analogous pattern to fixed delegate instructions produces similar gains in multi-step agentic pipelines. Method: Three pre-registered controlled experiments used Claude Haiku 4.5 delegates (n=5 per condition, temperature 0.5) assigned either a single-copy or a repeated-prompt instruction under blinded binary rubric scoring, totalling 30 sessions and 3,196 messages. Results: Experiment 1 (session-ID refactoring, 6 criteria) yielded a non-significant score delta of +0.30 with five of six criteria saturated at 100% in both groups (Fisher’s p=1.000). Experiment 2 (tree-sitter scanner evaluation, 7 criteria) produced a complete ceiling effect: all 10 runs scored 7/7 (Mann-Whitney U =12.5, p=1.000). Experiment 3 (Kotlin grammar synthesis, 7 criteria) revealed rubric-runner co-design failure: three of seven criteria scored 0/1 across both groups because the required investigations were absent from the runner prompt. On four reachable criteria, control scored a mean of 2.00/4 and treatment scored 2.40/4 (U =15, p=0.607). Across all three pilot experiments, we detect no effect of prompt repetition on task success. Treatment agents used 30.6% fewer total tokens in Exp1 but 7.2% and 19.9% more in Exp2–3; the direction reverses with session length, as long sessions save turns while short sessions pay pure overhead. Conclusion: Criteria requiring investigations absent from the runner prompt are unreachable by both groups and must be resolved before effect sizes are interpretable This is a preprint; it has not undergone peer review.