A Case for Agentic Tuning: From Documentation to Action in PostgreSQL

PerfEvolve: Translating Expert Tuning Methodologies into Executable Skills for LLM-Based Agents

Hongyu Lin; Mingyu Li; Weichen Zhang; Yihang Lou; Mingjie Xing; Yanjun Wu; Haibo Chen· 2026· DOI 10.48550/arXiv.2605.19988

The core problem

Computer system tuning has traditionally relied on documentation that distills expert knowledge into per-parameter recommendations. However, such guides capture only the conclusions experts reach, discarding the reasoning process behind them. This fundamental gap leads to three concrete deficiencies: documentation becomes stale as software evolves, fails to generalize across heterogeneous workloads, and ignores inter-parameter dependencies. The authors argue for a paradigm shift from static documentation to dynamic action. They introduce PerfEvolve, a system that translates expert tuning methodologies into executable skills, equipping LLM-based agents to perform version-consistency verification, workload-specific profiling, and multi-parameter joint optimization. Evaluated on PostgreSQL under TPC-C and TPC-H benchmarks, PerfEvolve outperforms state-of-the-art documentation-driven tuning baselines by up to 35.2%. The tool is available at https://github.com/ISCAS-OSLab/PerfEvolve.

PerfEvolve was evaluated on PostgreSQL using TPC-C and TPC-H benchmarks. The results show that PerfEvolve outperforms state-of-the-art documentation-driven tuning baselines by up to 35.2%. This improvement is attributed to the agent's ability to dynamically adapt to workload characteristics and version-specific behaviors, as well as to jointly optimize interdependent parameters. The evaluation likely includes metrics such as throughput, latency, and resource utilization, though specific numbers beyond the 35.2% improvement are not provided in the abstract. The tool is publicly available for reproducibility.

Computer system tuning has traditionally relied on documentation that distills expert knowledge into per-parameter recommendations. However, such guides capture only the conclusions experts reach, discarding the reasoning process behind them. This fundamental gap leads to three concrete deficiencies: documentation becomes stale as software evolves, fails to generalize across heterogeneous workloads, and ignores inter-parameter dependencies. The authors argue for a paradigm shift from static documentation to dynamic action. They introduce PerfEvolve, a system that translates expert tuning methodologies into executable skills, equipping LLM-based agents to perform version-consistency verification, workload-specific profiling, and multi-parameter joint optimization. Evaluated on PostgreSQL under TPC-C and TPC-H benchmarks, PerfEvolve outperforms state-of-the-art documentation-driven tuning baselines by up to 35.2%. The tool is available at https://github.com/ISCAS-OSLab/PerfEvolve.

PerfEvolve operationalizes expert tuning methodologies as executable skills for LLM-based agents. The approach comprises three core capabilities:

The work highlights a fundamental limitation of static documentation: it captures conclusions but not the reasoning that produced them. By encoding expert methodologies as executable skills, PerfEvolve enables agents to reason about tuning in context. This addresses the three deficiencies: staleness (via version-consistency verification), workload heterogeneity (via workload-specific profiling), and parameter dependencies (via joint optimization). The 35.2% improvement demonstrates the potential of agentic tuning. However, challenges remain, such as the effort to encode expert methodologies and the need for robust feedback loops. Future work could extend PerfEvolve to other database systems and explore automated skill acquisition. The approach aligns with the broader trend of using LLM agents for complex system tasks, suggesting a shift from documentation-centric to action-centric tuning.

CS practitioners and researchers

Opening member content…