OpenAI's practical guide to building AI agents is worth reading even if you are not building anything

OpenAI published a guide on building AI agents https://openai.com/business/guides-and-resources/a-practical-guide-to-building-ai-agents/ and it is more useful for non-builders than the title implies because it clarifies what agents actually are beyond the hype.

The most useful part is the section on when to use an agent versus when not to. Not every workflow benefits from an agent. Tasks that are well-defined, deterministic and short are often better handled by a simple automation or a direct API call. Agents are specifically valuable for tasks that require multi-step reasoning, judgment under ambiguity, or adapting to intermediate results in ways that a predefined script cannot handle.

The guardrails section is the part the marketing tends to skip. Building an agent without designing the constraints on what it can do and what it cannot touch is how production incidents happen. The guide covers this practically rather than as a theoretical safety concern.

The orchestration section, covering how agents hand off between each other and how the overall system maintains coherent state across multiple agent steps, is the architecture thinking that separates well-designed agent systems from brittle ones.

For anyone building on top of AI platforms: what is the most important constraint you have put on an agent to prevent it from doing something you did not intend?