Agents need control flow, not more prompts

The article argues that the current trend of relying on elaborate prompt chains to build AI agents is fundamentally flawed for complex, reliable systems. Prompts are non-deterministic, weakly specified, and difficult to verify, leading to unpredictable behavior and silent failures. The author points out that resorting to tricks like "MANDATORY" or "DO NOT SKIP" signals hitting the ceiling of prompting. In contrast, traditional software scales through recursive composability—systems built from libraries, modules, and functions that expose predictable behavior and enable local reasoning. The core thesis is that reliable agents need deterministic control flow encoded in software, treating the LLM as a component within a robust runtime, not as the system itself.

To achieve this, the article proposes deterministic scaffolds with explicit state transitions and validation checkpoints. But deterministic orchestration is only half the battle. In systems prone to silent failures, an agent without aggressive error detection quickly reaches wrong conclusions. Without programmatic verification, developers are left with three unsatisfactory options: the "Babysitter" (keeping a human in the loop), the "Auditor" (exhaustive end-to-end verification after the fact), or "Prayer" (blindly accepting outputs). The path forward is to encode logic in code, not prose, ensuring reliability through software engineering practices rather than prompt engineering.