arXiv:2610.00889v1 Announce Type: cross Abstract: Large language models (LLMs) have recently been applied in systems research as a tool to reduce human-intensive engineering effort through cost-efficient automation. Decades of research have produced a rich landscape of concurrency control (CC) protocols, each encoding distinct trade-offs in correctness, throughput, and abort behavior. However, mo
HakiCC is an LLM-driven multi-agent pipeline that automatically designs, verifies, and optimizes application-specific concurrency control (CC) protocols. Submitted on October 1, 2026, it addresses the need for expert knowledge in selecting CC protocols by automating the process for specific workloads.
The system operates in two stages: Stage 1 (Design & Verification): A multi-agent system generates a protocol implementation tailored to a workload description and iteratively repairs it until it passes conflict-serializability verification. Stage 2 (Optimization): An LLM-driven evolutionary loop refines the verified protocol to improve throughput while maintaining correctness.
In evaluations on TPC-C and AuctionMark benchmarks, HakiCC produced ten unique protocols. All were verified for correctness, and Stage 2 optimization yielded average throughput gains of +50.6% for TPC-C and +92.2% for AuctionMark. The generated protocols span diverse design families, including optimistic, pessimistic, versioned, and deterministic paradigms.
HakiCC: LLM-Driven Multi-Agent Design and Optimization of Concurrency Control Protocols
The paper presents HakiCC, an LLM-driven multi-agent framework for designing and optimizing concurrency control (CC) protocols in transactional systems. It frames CC design as a rich but historically human-intensive search space: existing protocols encode different assumptions about transaction ordering, versioning, locking or validation, abort behavior, and retry policies, each yielding distinct trade-offs among correctness, throughput, and stability under contention. Rather than treating CC protocol engineering as a purely manual systems-design task, the work uses LLMs to automate and structure the exploration of plausible protocol designs and refinements.
Technically, the key contribution is a multi-agent workflow in which LLMs are not simply asked to generate code, but are organized to collaborate on a systems-design problem. Agents can propose candidate CC mechanisms or policy choices, check logical consistency and correctness constraints, and reason about performance implications under workload and contention conditions. This makes the approach more disciplined than single-shot LLM generation: it couples protocol synthesis with validation and optimization, allowing the system to iteratively refine designs toward better correctness-throughput-abort trade-offs.
The material matters because it extends LLM-based automation from code completion or narrow task assistance to deeper systems engineering. Concurrency control is a core database and distributed-systems problem where small protocol choices can have large performance and reliability consequences, yet the design space is difficult to explore exhaustively by hand. HakiCC suggests a path toward adaptive, workload-aware protocol engineering, reducing the human cost of tuning database internals and potentially surfacing non-obvious protocol variants that would be expensive to discover through manual experimentation alone.