Agents can't choose between structure and flexibility

The debate around Python and Markdown looks like a choice between structure and flexibility, but pushing either extreme fails. Designing agents with code alone keeps paths stable but strips away reasoning and adaptation, while solving everything with natural language alone is flexible but impossible to control or debug.

A code-centric approach essentially ports a runbook directly into a workflow. It can work well for specific alerts, like an SRE would handle, but it easily breaks down when a new type of failure or a system change appears. It can't show people what was tried and what was ruled out, and the harness has to be reworked every time the environment changes.

Conversely, a Markdown-centric approach just gives a goal and lets the model figure it out. For one-shot tasks like generating slides, if the flow gets tangled or the layout breaks, there's no way to make fine-grained fixes—all that's left is repeatedly re-throwing prompts. In actual production environments, things like context management, model selection, cost control, and sub-agent orchestration must be handled by code.

The key isn't a compromise between the two, but a division of roles. Markdown should carry intent and domain knowledge, while code should handle execution, verification, enforcement, and tool calls. Only this way can agents behave like real agents—examining multiple hypotheses in parallel, letting humans check the reasoning process, and adapting to system changes.