Agents Need Control Flow, Not More Prompts

This is a solid anti-hype reminder: if the LLM owns routing and branching, you do not have an agent system so much as a brittle suggestion engine. The real production boundary is between probabilistic generation and deterministic execution.

• –Recursive composability comes from code, not prose; modules and functions are easier to reason about than stacked prompts.
• –Verification checkpoints matter more than prettier instructions, because silent failure is the default mode for agents.
• –The “babysitter / auditor / prayer” framing is useful: every agent system needs an explicit error-handling strategy.
• –This aligns with the broader move toward state machines, typed outputs, and tool policies instead of prompt gymnastics.
• –The practical takeaway: keep ambiguity inside the model, but keep control flow, retries, and validation outside it.