An Empirical Comparison of Multi-Agent LLM Architectural Patterns for Automated Unit Test Generation

Abstract

Software engineers are increasingly adopting multi-agent large language model (LLM) systems, yet the controlled empirical evidence comparing the underlying architectural patterns’ trade offs on software engineering tasks is still lacking. This thesis applies a two phase mixed-methods design. The first phase conducts a structured review of 31 articles published between 2023 and 2026, resulting in a taxonomy of 10 coordination patterns for multi-agent LLM systems. In the second phase, an empirical experiment comparing three patterns a Single-Agent Baseline, a Sequential architecture, and a Hierarchical architecture is carried out on automated unit test generation using the TestEval benchmark. Five dependent variables are evaluated (success rate, error handling, latency, cost, and agent communication turns) running each architecture independently on 30 tasks, all using Claude Sonnet 4, for 270 task executions in total. The Sequential pattern recorded the highest success rate (92.2%), the lowest latency variance (295.56 s2), and a competitive cost (0.089 USD per task). The Single-Agent Baseline reached a moderate success rate (81.1%) at the lowest cost (0.083 USD per task), while the Hierarchical pattern recorded the lowest success rate (54.4%), the highest median latency (142.37 s), and the highest mean cost (0.308 USD per task). The hierarchical pattern suffered from supervisor information bottleneck according to the qualitative results confirmed by 100% of the failing tasks of the architecture. These results come from a single LLM, a single benchmark, and non-optimized prompts per architecture, which means replication is needed before these findings are treated as general architectural principles.