Multi-agent LLM systems can organize work as symmetric peers or hub-and-spoke subagents, but controlled empirical evidence on when each topology produces better outcomes is scarce. We present three experiments comparing peer agents against subagents on software engineering tasks of increasing structural complexity: bug-fixing in a Rust codebase (n = 24), feature implementation in a Python library (n = 24), and software architecture design with LLM-simulated stakeholders holding asymmetrically distributed requirements (n = 32). The first two experiments yield ceiling effects-all treatments achieve perfect quality scores with zero peer-to-peer communication, confirming that code-level tasks without information asymmetry reduce multi-agent coordination to parallelism. The third experiment, which enforces information asymmetry by partitioning stakeholders across agents, produces the first significant result: peers outperform subagents with large effect sizes (composite d = +0.99, p = 0.014; resolution quality d = +1.04, p = 0.011). The treatment advantage for conflict resolution is approximately 3.5 times larger when 75% of conflicts span the partition boundary compared to 25%, supporting a directional dose-response pattern between cross-agent information dependency and output quality. Despite the peer messaging channel being available, agents never use it directly across any experiment; all inter-agent information flow is coordinator-mediated, suggesting the advantage stems from the peer topologylong-lived agents with shared file context and iterative relay-rather than explicit collaboration. Two independent evaluation methods (a rubric-based assessment and a blind architectural review) converge on the treatment advantage while measuring complementary qualities. Information asymmetry, not task difficulty, is the structural condition that determines whether multi-agent coordination improves output quality beyond parallelism.