The author contrasts the performance of DeepSeek V4 Flash with two different code agent runtimes. A local test showed Codex plus Flash successfully completed a complex, cross-file task. Conversely, Claude Code plus Flash initiated multiple reviews but their quality was unverified. This highlights that the effectiveness of an agent depends on the entire runtime, not just the model. The runtime encompasses the model, protocol, tools, context, recovery mechanisms, and acceptance criteria. Four key layers influence the outcome: protocol defines the interaction trajectory, tools act as structured contracts, context and recovery ensure task durability over time, and acceptance defines the completion state. DeepSeek's public updates suggest adaptations for specific models and environments, not universal dominance. The author proposes that agent effectiveness is a product of model potential multiplied by the harness's realization rate. A credible comparison requires fixing all runtime variables except the model. Otherwise, results should be considered runtime observations, not model leaderboards. The model sets the potential, but the runtime determines how much of that potential is realized. The author prompts reflection on the weakest link in current agent usage.