A new thesis argues that AI agents tackling complex tasks need deterministic control flow encoded in software, not increasingly elaborate prompt chains. The core problem: if you've ever resorted to MANDATORY or DO NOT SKIP in your prompts, you've hit the ceiling of what prompting can achieve.

Prompts lack the properties that make software reliable. They're non-deterministic, weakly specified, and difficult to verify. Code scales through recursive composability—libraries, modules, and functions built on each other—exposing predictable behavior that enables local reasoning. Prompt chains lack this property entirely.

The solution requires moving logic out of prose and into runtime. Developers need deterministic scaffolds: explicit state transitions and validation checkpoints that treat the LLM as a component, not the entire system. But orchestration alone isn't enough. Without aggressive error detection, an agent is simply a fast way to reach the wrong conclusion.

When programmatic verification isn't built in, teams face three choices: keep a human in the loop as a babysitter, perform exhaustive end-to-end auditing after each run, or simply vibe accept the outputs and hope for the best.