Agents can't choose between structure and flexibility (frontierai.substack.com)

The ongoing debate in the AI community between Python and Markdown as specification languages for agents underscores a critical architectural question. The Python camp emphasizes strict, code-based requirements, aiming for reliability in task execution, while the Markdown camp advocates for flexibility, allowing agents to explore and solve problems creatively. However, both extreme approaches are identified as failure modes, as they limit the agent's ability to adapt and reason effectively. Rather than accepting one side, experts argue for a hybrid approach: using Markdown for high-level intent and guidance coupled with Python for execution, enforcement, and reliable outcomes. This hybrid architecture is gaining traction among serious agent developers because it aligns with the core capabilities agents should have—reasoning flexibility and deterministic guardrails. By navigating complex tasks, this framework allows agents to evaluate multiple hypotheses, provide human-readable insights into their reasoning process, and adapt to changing conditions without requiring manual updates to their operational structure. Thus, the real focus is on understanding the unique needs of each component within an agent's operation and intentionally designing systems that capitalize on both flexibility and control, moving beyond the simplistic dichotomy of Python versus Markdown.