Preprint finds adapting agent instructions improves scores across seven models

Researchers report in a preprint that adapting reusable instructions to an AI agent's capabilities improved final benchmark scores by 2.0-5.1 points across seven models, compared with the same iterative improvement process without adaptation guidance. The approach changes instructions rather than model parameters. The researchers diagnosed failures in tool use, context management, instruction following and verification, then turned those diagnoses into guidance for rewriting instructions. Each model revised its instructions over five rounds using execution records and verifier feedback. The gains address a mismatch: across 103 task-skill pairs, instructions that improved GPT-5.6-luna by 31.1 points degraded Gemma-4-E4B by 8.8 points. The evaluation covered three benchmarks.