Agents Need Control Flow, Not More Prompts

The Prompting Ceiling

If you have ever written MANDATORY or DO NOT SKIP in an AI agent prompt, you have hit the ceiling of prompt-based approaches. Developer Bryan Suh argues in a widely-shared HN post that reliable AI agents require deterministic control flow, not more elaborate prompting.

Why Prompt Chains Fall Short

Suh frames LLMs as a programming language where statements are suggestions and functions return Success while hallucinating. In this environment, predictable behavior and local reasoning become nearly impossible. Prompt chains are non-deterministic, weakly specified, and difficult to verify.

Traditional software scales through recursive composability: libraries, modules, and functions stacked reliably together. Prompt chains lack this property entirely.

The Solution: Deterministic Scaffolds

Rather than treating the LLM as the entire system, it should be a component within a larger architecture with explicit state transitions and validation checkpoints. Logic must move out of prose and into runtime. The LLM handles ambiguity and natural language; the scaffold handles correctness and flow.

The Error Detection Problem

Even deterministic orchestration is insufficient without aggressive error detection. Without it, an agent becomes a fast way to reach the wrong conclusion. Constant human oversight, exhaustive post-run verification, or accepting outputs without verification — none of these scale.

Takeaway

The post earned 552 upvotes on HN. For anyone building agents, the message is clear: architectural rigor, not prompt elaboration, is what makes complex agent systems reliable.