Ahead-of-Time (AOT) agent harnesses suffer from fundamental precompilation and uniformity bottlenecks when confronted with diverse, real-world tasks. Under an AOT paradigm, the runtime scaffold is treated as a durable, static artifact optimized over prior experience under the assumption that a single precompiled configuration will generalize across future domains. This approach creates critical limitations:

• Precompilation Constraint: Optimization loops are forced to synthesize memory layouts, planning strategies, and tool orchestrations before observing the specific structure or failure surfaces of an incoming problem instance.
• The Uniformity Fallacy: Imposing a uniform, static scaffold across heterogeneous roles—analogous to forcing an accountant, a graphic designer, and a janitor to work with identical desk configurations and software suites—fails because operational requirements are instance-dependent rather than universal.
• Intractable Design Search: Different tasks require mutually incompatible harness priors, such as parallel evidence exploration for wide search, lean serial loops for terminal environments, or dedicated working memory for deep research. Searching a massive design space to locate a single compromise scaffold is cumbersome and fundamentally incapable of matching heterogeneous instance requirements.