A software engineer with deep expertise in programming has raised concerns about the safety and alignment of AI agents, particularly in areas beyond their creators' expertise. While developers may excel in specific domains, they often depend heavily on the underlying model's priors to handle other complex concerns, which introduces significant unknown risks. The engineer points out that AI models frequently produce output that, although functional, contains flaws or inefficiencies—referred to as "slop"—which experts recognize as problematic. These issues arise because models are trained and rewarded based on behaviors that may not align with expert standards, leading to compounded misalignments over time. Furthermore, AI systems lack mechanisms to anticipate future regret or maintain long-term coherence, making sustained, reliable agentic behavior difficult to achieve. The engineer also highlights the challenge of defining what constitutes acceptable shortcuts in AI decision-making, as interpretations vary widely depending on individual values and goals. This complexity underscores that achieving true alignment in AI is an inherently difficult problem without straightforward solutions. The insights emphasize the need for cautious development and evaluation of AI agents, especially when tasked with high-stakes objectives.