Reddit debate questions agentic workflows: probabilistic flaws and brittleness

The viral Reddit thread scrutinizes agentic workflows—multi-step AI automation where large language models make decisions at each stage. The author points out that unlike deterministic RPA, which follows rigid rules, agentic systems introduce probabilistic reasoning (e.g., GPT-4o’s 5-10% error rate per step). When chaining five such steps, the chance of a flawless outcome drops below 60%, making the overall process unreliable for critical tasks. This randomness, the post argues, forces humans to audit every action, defeating the purpose of automation.

Another core criticism is brittleness. Agentic workflows often rely on scraping websites or interacting with APIs that can change without notice—a website redesign, an updated plugin requirement, or altered data formats. Unlike adaptive human workers, these systems break silently. The author also raises concerns about the web's increasing hostility to bots (via CAPTCHAs, anti-scraping measures) and the lack of explainability in AI-driven decisions, which hinders debugging and compliance. They conclude that agentic workflows are only viable for extremely narrow, stable use cases, similar to existing RPA, and that the hype may be overblown.