AI Agent Reliability Dilemma: What is Really Needed is Control Flow, Not More Prompt Words

Core conclusion

When the AI Agent needs to handle complex multi-step tasks, stacking longer prompt words does not solve the fundamental problem. On May 7, an article titled "Agents need control flow, not more prompts" scored 519 points on Hacker News, triggering Silicon Valley engineers to deeply reflect on the Agent architecture.

Key Points

• Time of incident: 2026-05-07
• Core idea: A reliable Agent system requires deterministic control flow (code) rather than non-deterministic prompt word chains
• Impacts: All content teams and developers who use LLM Agent to perform automated tasks
• Derived frameworks: LangGraph, BAML, Stripe Minions, SalesForce AgentScript and other control flow solutions

Background: When the prompt word meets the ceiling

If you have ever written MANDATORY or DO NOT SKIP in a prompt word, you have hit the prompt word ceiling.

The author bsuh points out a poignant metaphor: imagine a programming language in which every statement is just a "suggestion" and functions returning "success" may still be hallucinating. Reasoning becomes impossible, and as complexity grows, reliability breaks down completely.

The foundation of software engineering lies in recursive composability: systems are built layer by layer of libraries, modules, and functions. Code exposes predictable behavior and supports local reasoning. But hint word chains do not have this property - they are non-deterministic, weakly specified, and difficult to verify.

Key Impact

Three solutions to unreliable agents

bsuh draws a line in the article: deterministic orchestration is only half the problem. In systems prone to silent failures, agents without active error detection simply arrive at the wrong conclusion faster. Without programmatic verification, developers are left with only three options:

1. Babysitter (nanny mode)

Keep humans in the loop and intercept errors before they spread. Suitable for scenarios with high reliability requirements, but cannot be scaled up.

2. Auditor (audit mode)

Do a thorough end-to-end verification after the run. Good for batch processing tasks, but you won’t notice problems until it’s too late.

3. Prayer (prayer mode)

"Vibe accept the outputs" - accept the outputs and pray. This is the true state of most teams right now.

Community Voice: Practice of Control Flow Scheme

HN community discussion furthers this idea:

• apalmer pointed out: The breakthrough in AI coding lies not in the improvement of AI intelligence, but in the movement of core process execution from LLM prompt words to the framework layer.
• Stripe's Minions Team Share: Inserting deterministic nodes between non-deterministic LLM jobs to handle quality assurance, leaving LLMs to just do what they do best.
• 59nadir mentioned that Stripe's Minions system is exactly this idea in action - embedding deterministic QA nodes in LLM workflows.
• jerf Summary: In scenarios where prompt words are unreliable, what is needed is "next generation AI" - not a stronger LLM, but a better architecture.

Tool entry

The following tools/frameworks appear in the text: LangGraph, LangSmith, Claude, ChatGPT, Stripe, SalesForce, BAML, n8n

Internal link guidance

• Want to learn more about the practical use of AI Agent tools? Watch: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes
• Real automated monetization case: He earns over 10,000 per month by relying on AI code review + specification-driven development: a practical review of a freelance developer